물리적 거래 전략

실제 상품 거래 전략.

A 상품 무역 회사의 물리적 자산 강도 40. 이러한 자금 조달 전략은 상품 거래소에서 거래 될 수 있고 농산물을 포함 할 수있는 물리적 상품 유형과 일치한다.

곡물 및 오일 시드 선물을 통한 헤징에 대한 자율 학습 안내서. 거래 선물 및 옵션 시장. 또는 실제 물리적 상품의 인도를 통해

균열 확산 차트 2014.

상품 정보 시스템.

주식 시장에서 ppt로 상품 주식 브로커가 되십시오.

홈 서비스 상품.

상품 옵션 온라인 거래.

선물 계약은 표준화 된 단위로 매우 가시적으로 거래됩니다. 물리적 상품 계약은 실제 배달을 허용, 그들은 모든 시장 참여자를 보장합니다. 선물 거래의 세계에서 실제 세계에 초점을 상품 무역 회사 파생 상품 스탠드 비행기입니다. 온라인 골드 무역 bforex는 단순 온라인 금 거래를 제공합니다. 미국 상품 선물 거래. 단기적으로 적극적인 투자 전략을 추구합니다.

ENR TAX 상품 무역 회사 중요한 성공 요인으로 무역 집중 2012 년 10 월 kpmg KPMG INTERNATIONAL. PhysOps Visual Cockpit은 상품 거래를 간소화하는 데 도움이되는 상품 거래, 일정 관리 및 물류 소프트웨어 솔루션입니다.

LinkedIn에 681 물리적 거래 기회가 있습니다.

상품 거래 시스템.

상품 선물 옵션 거래.

칼리 어 (Kallier)는 개인 소유의 물리적 상품 거래, 중개 및 상품 자문, 비철금속 및 비철금속 및 미네랄 거래입니다.

401k 롤오버 옵션.

상품 시장은 현물 가격을 이용한 물리적 거래 및 파생 상품 거래, 투자자가 상품 거래의 기초를 배우는 것을 돕기 위해 제공되는 교육 자원을 포함 할 수 있습니다.

이진 옵션 전략.

상품 헤징.

상품 중개인.

직업 선물.

Ellefsen 2010 년 5 월 13 일 부분적으로 기계 공학부에 제출되었습니다.

고급 석유 거래 전략 : 물리 및 파생 코스 (두바이, 아랍 에미리트 - 2017 년 11 월 14-16 일) - 연구 및 시장.

이 과정은 실제 거래와 파생 거래의 경험이있는 거래 / 교육 전문가가 진행하는 양방향 팀 기반 거래 과정으로 개발되었습니다. 이 과정은 거래 시뮬레이션이 아니며 거래 기법을 설명하고 대표와 공유하는 독특한 학습 경험입니다.

처음에는 팀이 무역 및 관리를위한 물리적 석유 직책을 받게됩니다. 팀은 시장 데이터를 분석하고 거래 의사 결정을하도록 장려됩니다. 각 팀은 거래하는 제품 (원유 또는 정제 된 제품)에 따라 다양한 문제를 갖게됩니다.

의사 결정자 (코스 강사)가 피드백을 제공하고 시장 정보를 가장 잘 적용하는 방법을 교육하는 대화 형 환경에서 의사 결정을 검토하고 논의합니다. 팀은 헤지 할 가장 좋은 도구 인 헤징시기와 실제 거래에서 누려야하는 위험 요소를 고려하여 실제 위치를 헤지하는 방법을 살펴볼 것입니다.

결정은 3 일 기간 동안 시장 상황이 변함에 따라 지속적으로 검토되며, 이 시점에는 포지션이 종결되고 헤지가 풀립니다. 또한 팀이 파생 상품 거래를 선택하고 수행하는 과정에서 파생 상품 거래를 관리하고 관리 할 수있는 기회도 제공됩니다.

과정 참여 점심 도시락 및 커피 휴식 프로그램 하드 카피 및 USB 장치의 과정 자료 산업 인증서 비자 지원 (필요한 경우)

누구나 석유 거래에 관심이 있지만, 석유 시장과 석유 파생 시장에 대한 기본적인 지식을 가진 사람들.

상품 시장 모델링 및 물리적 거래 전략 -.

문서 세부 정보를 편집하려면 누릅니다.

포맷되지 않은 텍스트 미리보기 : 당 나르 S. Ellefsen 엔지니어의 드 난 에콜 폴리 테크닉 대학의에서 기계 공학에서 과학 석사의 학위 요건의 부분적인 성취 기계 공학학과에 제출 한 2008 년 상품 시장 모델링과 실제 거래 전략 매사추세츠 공과 대학교 2010 년 6 월 © 2010 Massachusetts Institute of Technology. 판권 소유. 저자의 서명 : _____________________________________________________________ 기계 공학과 2010년 5월 13일에 의해 인증 : ____________________________________________________________________에 의해 허용 기계 공학 및 조선 논문 감독자의 폴 D. Sclavounos 교수 : ___________________________________________________________________ 대학원생 2 기계 공학 회장의 데이비드 E. 하트 교수, 학과위원회 Einar S. Ellefsen의 상품 시장 모델링 및 물리적 거래 전략 기계 공학 석사 학위 요건을 부분적으로 이행하기 위해 2010 년 5 월 13 일에 기계 공학부에 제출 ABSTRACT 상품과 관련된 투자 및 운영 결정이 내려집니다. 이 상품의 선물 가격에 근거하여 이러한 가격은 변동성이 크며, 진화 모델은 변동성 및 상관 관계 기간 구조를 정확히 설명해야합니다. 순방향 곡선의 2 요소 모델을 제안하고 원유, 운송, 천연 가스 및 난방유 시장에 맞게 조정합니다. 이 모델의 이론적 인 특성은 앞으로의 곡선의 레벨과 기울기에 영향을 미치는 독립적 인 요인으로의 분해에 초점을 맞추어 조사됩니다. 두 요소 모델은 상품 가격과 관련된 두 가지 문제에 적용됩니다. 아시아 옵션의 가격에 대한 대략적인 분석식이 유도되어 선적 옵션의 시장 가격을 설명합니다. 2008 년 말 석유 시장에 등장한 부유 무역은 최적의 정지 문제로 제시됩니다. 순방향 곡선의 2 요소 모델을 사용하여 원유 저장 가치가 역사적으로 도출되고 분석됩니다. 개발 된 물리적 상품 거래를위한 분석적 틀은 예상 이익, 관련 위험 및 주요 위험 요인에 대한 노출을 계산할 수있게합니다. 이것은 시장 참여자가 그러한 물리적 거래를 미리 분석하고, 화물을 언제 판매 할 것인지에 대한 결정 규칙을 만들고, 앞으로 커브에 대한 노출을 정확하게 헤지하도록 허용합니다. 논문 감독자 : Paul D. Sclavounos 교수 : 기계 공학 및 해군 건축학 교수 3 4 목차 1. 서론. 7 1. 상품 시장. 7 2. 정의 및 시장. 3. 동기 부여. 10 4. 목표. 5. 방법론 및 개요. 11 2. 시장 모델링. 12 1. 이론적 근거. .12 2. 기존 문헌. 3. 탐색 적 데이터 분석. 13 4. 상품 선물의 2 요소 모델. ... 15 5. 주요 구성 요소 분석. 6. 앞으로의 곡선 계절성. ... 21 7. 마켓 조정. 8. 리스크 중립에서 객관적 인 측정으로 전향 위험 프리미엄. 29 9. 세 가지 요소로 확장. 정적 정방향 커브의 모델. 31 11. 시장 모델의 응용. 33 3. 상품에 아시아 옵션. 37 1. 정의 및 시장. ... 37 2. 아시아 옵션에 관한 서적. 38 3. 2 요소 모델에서의 대략적인 수식. 38 4. 다른 아시아 옵션 모델 및 시장 가격과의 비교. 46 5. 아시아 옵션의 헤지. 53 6. 아시아 옵션 가격의 매개 변수 의존성. 56 5 4. 부유물 저장 무역. 58 1. 서론. 58 2. 부동 스토리지 문제. 58 3. 솔루션 방법. ... 62 4. 분석 특성. 5. 이익과 위험. 68 6. ​​결과. 70 7. 초과 이익의 기원. 90 8. 일반 상품 거래 문제. 95 5. 결론. 98 1. 결과 요약. 98 2. 향후 연구를위한 제안. 99 6. 부록. 100 2. 스팟 가격 과정은 2 요소 모델에 의해 암시된다. 101 3. 주요 요소 두 요소 모델의 분석. 4. 2 요인 모형 하에서 일정한 성숙 곡선의 진화. 103 5. 일정한 성숙도의 순방향 곡선에 대한 세 번째 요인의 영향. 6. 평균 가격 계약의 블랙 변동성. 106 7. 최적의 정지 문제에 대한 준 - 분석 솔루션. 108 8. 상품 파생 상품 시장 거래량. 플로팅 스토리지 거래에 사용되는 100 개의 루트, 화물 및 배송. 참조. 1. 서론 1. 상품 시장 2008 년 7 월 3 일 브렌트 원유 선물은 배럴당 146 달러에 거래되었다. TD3 아라비안 걸프 - 일본 운송 경로는 240 Worldscale에서 인용되었으며 분석가들은 다음 달 안에 원유 가격을 200 달러로 예측했다. 12 월 3 일, 북해산 브렌트 유는 배럴당 45 달러, TD3는 70 %로 각각 69 %, 71 % 하락했다. 상품 시장은 세계에서 가장 변동성이 크며, 변동성은 관련된 행위자의 이익과 위험의 원천입니다. 이러한 위험을 관리하기 위해 실제 현물 시장은 초기 시장에서 동 시장을 동반하고 금융 파생 상품 시장으로 전환되었습니다. 시카고 무역위원회는 1848 년에 농산물에 대한 외환 거래 계약을 도입했으며, 1860 년대 초반부터 원유를 거래했습니다 (Yergin, 2008). 가장 현대적인 상품 시장은 물리적 시장과 금융 시장이라는 두 개의 고리로 연결된 시장으로 구성됩니다. 물리적 또는 현물 시장은 상품을 판매하거나 배달하는 모든 시장 참여자들로 구성됩니다. 원유 시장에서 이들은 석유 회사, 정유 회사 및 물리적 무역 회사입니다. 현물 시장에서의 거래는 일반적으로 특정 날짜와 위치에있는 브로커, 일치하는 판매자 및 구매자를 통해 발생합니다. 금융 상품 시장은 그 지점을 기반으로하는 파생 상품 시장입니다. 이러한 파생 상품은 선도, 선물 및 옵션의 형태를 취하며 물리적 시장에 관여하는 회사와 다른 플레이어의 추측에 의한 위험 관리에 사용됩니다. 중요한 것은 파생 상품이 물리적 시장에 대항하여 정착함으로써 두 가지를 연결한다는 것입니다. 경우에 따라 파생 상품이 물리적으로 결제됩니다. 즉 구매자가 실제 상품을 수령합니다. 다른 기업의 경우, 파생 상품은 물리적 시장에서의 거래를 토대로 매일 공표 된 스팟 지수와 재정적으로 합의합니다. 금융 시장과 물리적 시장의 상대적 규모는 파생 상품 시장의 발전 수준에 달려 있습니다. 부록 1에서 볼 수 있듯이, 2009 년에 원유로 거래되는 파생 상품 (선물 및 옵션)의 양은 3,310 억 배럴이었으며 (CIA, 2009), 이는 파생 상품 시장을 9 배 실제 시장의 크기. 유조선 운송에서 파생 상품 시장은 2009 년에 3 억 4 천만 톤의 석유화물을 거래했으며, 2006 년에 거래 된 1 억 4,500 만개의 유조선 재량 (스톱 포드, 2009)과 비교하여 파생 상품 시장 규모를 실제 시장 규모의 두 배로 평가합니다. 1980 년대 이후 석유 시장에서 일어났던 것과 유사하게 물가 조정을위한 표준화 및 협약을 통해 이루어지는화물 파생 상품 시장에서 여전히 큰 성장 잠재력이 존재합니다. 상품에 대한 현물 시장과 선도 시장 간의 연계성이이 논문의 주요 주제이며, 특히 금융 시장이 어떻게 실제 거래 결정에 대한 더 큰 통찰력을 얻는 데 사용될 수 있는지에 관한 것이다. 원유 및 유조선 운송에 중점을두고 있지만 석탄과 같은 건어물에는 동일한 원칙이 적용됩니다. 정의와 시장 정의이 논문에서 우리는 상품이 날짜 t에 현물 가격 S (t)로 거래되는 상품 시장을 고려할 것이다. 가능한 한 빨리 납품을위한 시장 가격으로 전기 다음날 또는 원유에 대한 다음 달이 될 수 있습니다. 이 시장과 관련하여 날짜 t에 대한 순 가격 F (t, T)가 있습니다. 이것들은 미래의 T 일자에 상품을 배달하기위한 시장 가격입니다. 그 차이는 앞으로의 계약과 미래의 계약 사이에서 종종 발생하며, 후자의 계약은 매일 더 표준화되고 시장에 진입합니다. 그러나 여기서 우리는 그러한 구분을하지 않을 것입니다. 선수금이 재정적으로 결제되는 경우, 결제일 T에 입력 한 선물 계약의 긴 포지션은 결제일 T에 S (T) - F (t, T)를 지불하게됩니다. 물리적 시장이 유동적이면, 따라서 물리적 또는 금융 선도 계약을 체결하는 것은 시장 위험과 동등합니다. 둘 다 F (t, T)의 고정 구매 가격을 제공합니다. 시장에서 거래되는 선물 거래는 순방향 곡선 F (t, T)를 구성하게합니다. 일반적으로 만기 T는 월간이지만 단기적으로 더 세분화 될 수 있습니다. 우리는 또한이 곡선을 만기일 τ = T - t로 지수화할 것인데, 이는 f (t, τ) = F (t, t + τ)와 같은 미래 계약의 정산 시점이다. 원유 시장 원유 파생 상품 시장은 2009 년에 3,030 억 배럴 (ICE, 2009 및 CME, 2009)으로 거래되는 가장 큰 원자재 시장입니다. 그러나 원유의 가격은 그 등급 (주로 비중과 황 함량)과 위치에 따라 다릅니다. BFOE (브렌트, 포티, Oseberg 및 Ekofisk) 북해와 경질유 쿠싱에서, 오클라호마에있는 미국에서, 또한 미국 서부 텍사스 산 중질유 (WTI)로 알려진 : 원유의 두 개의 기준 등급 그러나이있다. 다른 지역의 대부분의 등급의 오일은이 마커 원유와 다른 가격으로 책정됩니다. WTI 선물 계약은 뉴욕 상품 거래소 (NYMEX)에서 거래되며 계약 월 동안 오클라호마 쿠싱의 파이프 라인 시스템으로 물리적으로 인도 가능합니다. 이로 인해 WTI는 현 위치에서 원유 재고 계약을 체결하게된다. 그러나 쿠싱은 내륙에 위치하고 파이프 라인을 통해서만 접근 할 수 있지만 수입 원유는 일반적으로 멕시코만의 루이지애나 근해 석유 항구 (LOOP)에서 유조선으로 도착합니다. 따라서 우리는 Louisiana Light Sweet (LLS)의 현물 가격도 고려할 것입니다. BFOE는 가장 완벽한 원유 선물 시장입니다. 데이트 브렌트 (Brent Brent)로 알려진 현물 가격은 "플랫 츠 거래 창구 (Platts trading window)"기간 동안 플랫 츠 (Platts)에 의해 매일 평가되며 앞으로 10 일에서 21 일 사이에 배달되는화물에 해당합니다. 한 달 전부터 ICE (Intercontinental Exchange)에서 거래되고 ICE Brent Index에 대해 재정적으로 정착하는 선물 계약이 있습니다. 이 두 거래자는 Brent Contracts For Difference (CFD) 카운터를 사용하여 특정 시간대의 가격을 헤지 할 수 있습니다. 8 이러한 다양한 가격을 사용하여 BFOE에 대한 매우 정확한 전진 곡선을 작성할 수 있습니다 (특히 짧은 쪽에서). 선물 계약 이외에, 원유 마커 가격에 대한 옵션을위한 유동성 시장이 있습니다. 거래소 옵션은 대부분 미국식이며 행사시 선물 계약을 체결합니다. 유조선 운송 시장 석유는 1861 년부터 선박으로 운송되어 왔지만 (Yergin, 2008), 석유 메이저에 의해 통제 된 물류 운동에서 현물 시장으로의 전환은 상대적으로 최근에 이루어졌으며 1990 년대에 70 %의 전세가 20 % 현물 시장에서 유조선은 중개인을 통한 단일 항해 (예 : Sullom Voe-LOOP)를 위해 공인되었으며 모든 비용은 가격에 포함됩니다. TD1에서 TD18 (더러운, 즉 원유) 및 TC1에서 TC11 (청소, 즉 제품)으로 번호가 매겨진 더럽고 깨끗한 탱커에 대한 여러 참조 경로가 있습니다. 각 거래일 말에 발틱 교환소는 브로커를 조사하고 발트 익스프레스의 더럽고 깨끗한 유조선 지수를 구성하여 각 노선의 가격 수준에 대한 평가를 게시합니다. 이것은 유조선 시장에서 인정 된 현물 가격입니다. 유조선 요금은 일반적으로 Worldscale (WS)이라는 단위로 게시됩니다. 이 노선은 각 노선에 따라 Worldscale Association에 의해 매년 갱신되며 특정 노선에 대한 참고 탱커의 기준 가격을 재화 중량 당 미국 달러로 나타냅니다. WS100의 현물 가격은이 가격의 100 %와 같지만 WS150은이 가격의 150 %입니다. 항해 전세 용 유조선 요금에는 모든 비용, 즉 연료, 항구 및 운하 비용이 포함됩니다. 이 비용을 순 배수로 지불 한 일일 가격에 해당하는 배 당 TCE (TimeCharter Equivalent) 가격을 일일 미국 달러로 환원하는 것이 유용합니다. 이를 위해서는 거리, 선속, 연료 소비 및 연료 가격에 대한 세부 정보가 필요합니다. 발틱 교환국은이 평가를 바탕으로 VLCC, Suezmax, Aframax 및 MR 탱커의 TCE 가격을 매일 발표합니다. 유조선 시장은 FFA (Forward Freight Agreement) 시장이 상대적으로 최근에 개발 된 것을 보았습니다. BIFFEX1 선물은 1985 년에 일찍 거래되었지만 인기를 잃었고 이후 특정 FFAs로 대체되었습니다. FFAs는 계약 월말에 해당 월 중 발틱 익스체인지 지수의 일일 평균 가치 산술 평균치를 재정적으로 결정합니다. FFAs는 유조선 FFAs에서 가장 큰 시장 점유율을 가진 International Maritime Exchange (Imarex)와 함께 브로커를 통해 거래됩니다. 유동성은 TD3 (VLCC 아라비안 걸프 - 일본), TD5 (Suezmax West Africa - 미국 대서양 연안) 및 TC2 (MR 제품 유조선 로테르담 - 뉴욕)와 같은 몇 가지 주요 경로에 집중되어 있습니다. 브로커 교환소는 브로커 가격 외에도 폴링 브로커가 수집 한 FFAs를 매일 평가합니다. 해운 시장의 중요한 특이성은 무역되는 상품 (tonne-mile)이 서비스가 아니라 저장 될 수있는 물리적 상품임을 의미합니다. 이 점에서 전기 시장과 비슷합니다. 톤 마일을 저장하는 것이 불가능하지는 않지만 예를 들어 느린 김을 내거나 배를 세우는 일은 어렵습니다. 재고가 없기 때문에 현물 가격 변동성이 높아지고 다른 테너의 선급 계약 간의 상관 관계가 낮아집니다. 또한 Imarex가 앞장서고있는화물 옵션에 초기 시장이 있습니다. 이들은 아시아 스타일이며, FFAs와 마찬가지로 한 달 동안 현물 지수 평균에 정착합니다. 1 발틱 지수 1985 년부터 2001 년까지 발트해 연안 이니셔티브 인 Freight Futures Exchange가 존재했다. 3. 동기 부여 상품 가격의 변동성은 시장 행위자에게 상당한 위험을 노출시킨다. 상품과 관련된 모든 투자 결정은 투자자를 전방 곡선에 노출시킵니다. 이러한 결정에는 석탄 광산 구입, 발전소 운영, 새로운 선박의 주문 및 취소, 다른 지역 간 상품 거래, 상품에 대한 옵션 작성 등이 포함됩니다. Dixit and Pindyck (1994)에서 강조한 바와 같이, 이러한 결정은 가격 예측에만 근거해서는 안된다. 원유 현물 가격의 40 % 연간 변동성은 2 %의 예상 성장률보다 원유 자산 투자 결정에 더 큰 영향을 미칩니다. 2008 년은 석유 시장에서 특히 변동이 심한 해였습니다. 현물 가격이 급락하면서 수년간의 후퇴 이후에 가파른 contango 로의 전방 시장 전환으로도 나타났습니다. 동시에 유조선 요금은 71 % 하락했습니다. 이로 인해 유조선 배열이 부유 저장 시설로 사용되었고, 1973 년 이후로는 사용되지 않았던 현상 인 미사용 원유를 저장하고 운반하기 위해 운반 항구 근처에 정박했다. 또한 시장 및 가격 변동성에 따라 달라집니다. 많은 투자와 거래가 진행되고 있지만, 일반적으로 적절한 틀을 사용하여 평가되는 것은 아닙니다. 그러한 투자 나 거래에 대한 정확한 평가와 운영 의사 결정은 관련된 전방 곡선에 대해 간단하고 정확한 모델을 사용해야합니다. 이러한 모델은 회사의 리스크를 정확하게 관리하고 다양한 가능성에 대해 정보에 근거한 선택을 할 수있는 가능성을 열어줍니다. 4. 목표이 논문의 주된 목적은 상품의 최적의 물리적 거래를위한 간단하고 효율적인 틀을 개발하는 것이다. 그러한 틀은 우리가 2008 년 말에 등장하고 2009 년까지 계속 된 부 동차 무역을 이해하고 분석 할 수있게 해줄 것입니다. 이 논문에서 우리가 대답하려고하는 질문은 다음과 같습니다 : • 단순한 2 단계 전방 곡선 모델이 역사적 변동성 다른 상품 시장에서 거래 된 선도 계약의 상관 관계 • 그러한 2 요소 모델에 대한 현물 가격 과정의 결과는 무엇입니까? • 선적 시장에서 거래되는 상품 아시아 옵션은 시장 참여자에 의해 어떻게 해석되고 가격이 책정되고 헤지되어야합니까? 그러한 옵션에 대한 묵시적인 변동성의 의미는 무엇입니까? • 대서양 횡단 원유 차익 거래 창이 언제 열렸습니까? 이 무역에서 떠내려는 스토리지 기회는 언제 있었습니까? 10 • 상인의 이익을 극대화하기 위해 따라야 할 최적의 부유 스토리지 전략은 무엇입니까? 화물을 즉시 판매하지 않음으로써 포워드 곡선에 대한 노출을 유지하는 것이 가치가 있으며, 이 값을 어떻게 이해할 수 있습니까? • 일반적인 물적 거래 문제를 고려할 때 따라야 할 최적의 선박 경로 지정 전략은 무엇입니까? 선박을 언제 초기 목적지에서 다시 연결해야합니까? 5. 방법론과 개요 우리가 작업하게 될 일반적인 틀은 Musiela와 Rutkowski (2008)에서 공식화 된 이토의 확률 적 미적분학을 사용하는 연속적인 금융 시장의 기본 틀이다. 유가 증권 가격은 일반적으로 dS (t) = μ (t) dt + σ (t) dW (t) S (t)의 확산을 따르는 것으로 가정 할 것이다. 즉, 순간 이동과 σ (t)는 확률 과정 ​​S (t)의 순간적인 변동성이라고 부른다. 이 논문은 이론적이고 실용적입니다. 우리는 새로운 모델과 새로운 이론 결과를 제시합니다. 그러나 새로운 모델이나 결과를 발표 할 때마다 시장 데이터 또는 과거 실적에 대한 조정을 제시하고 그 결과를 분석 할 것입니다. 파트 2에서는 상품 순방향 곡선의 2 요소 모델을 제시하고 4 가지 다른 상품의 순방향 곡선의 주요 역사적 특징을 재현한다는 것을 보여줍니다. 우리는 또한 그것의 이론적 인 성질을 탐구하고 그것들을 일정한 성숙의 전진 곡선에 충격을주는 평균 복귀 요인의 관점에서 재구성한다. 이 곡선의 파라 메트릭 확률 모델을 사용하여 Part 3에서 FFAs와 같은 평균 가격 계약의 대략적인 진화를 유도 할 수 있습니다. 그러면 아시아 주요 옵션을 고려한 대략적인 형태의 닫힌 형태의 수식을 찾을 수 있습니다 상품 선물의 개념 : 상대적으로 짧은 평균 기간뿐만 아니라 가격 구조와 변동성의 기간 구조. 그런 다음 얻은 가격을 운송 옵션의 시장 가격과 비교하여 시장 데이터에 매우 적합하게 찾습니다. 제 4 부에서는 원유 순방향 곡선 모델, 해운 시장에 대한 데이터 및 확률 론적 동적 ​​프로그래밍을 사용하여 최적의 라우팅 및 부동 스토리지 문제를 공식화합니다. 대서양을 가로 지르는 원유 거래를위한 최적의 정지 문제를 공식화 한 후 2007-2009 년 동안이 무역의 경험적 결과를 조사하고 주요 특징을 확인합니다 : 그것이 잘 수행되고 어떤 기원 이익의. 11 2. 시장 모델링 1. 이론적 근거 미디어 및 상품 시장 분석가들은 예상되는 공급 및 수요 발전을 기반으로 가격 추세에 집중하는 경향이있다. 이는 중요한 일이지만 상품 시장은 변동성이 크며 원유와 마찬가지로 40 %의 가격 변동성으로 인해 2 %의 예상 성장률이 저하 될 것입니다. 미래의 수급 균형에 대한 시장의 기대는 다른 상품에 대한 선물 시장에 반영됩니다. 대부분의 상품 시장은 이제 만기가 긴 액체 순방향 곡선을 가지며 이러한 완벽한 순방향 곡선은 현물 가격뿐만 아니라 장기 투자 및 운영 결정을 이끌어 내야합니다. 이를 염두에두고, 원자재 가격 모델은 완전한 계약 곡선의 진화와 다양한 계약의 변동성에 대한 현실적인 모델을 제공해야합니다. 그런 모델은 운영상의 유연성을 가진 다른 파생 상품이나 실제 자산에 대한 가격 책정과 같은 다양한 애플리케이션에 사용될 수 있습니다. 이러한 모델은 또한 매개 변수의 수가 적어야하며 시장 가격에 맞게 조정될 때 현실에 해당해야합니다. 현실적인 파라 메트릭 모델을 사용하면 파트 3 및 4에 나와있는 것처럼 옵션 가격 및 실제 자산 가격에 대한 분석식을 쉽게 얻을 수 있습니다. 2. 기존 문헌 상품 시장에 대한 초기 연구는 현물 가격 모델링에 중점을 두었습니다. 유일한 관찰 가능한 시장 가격되었습니다. 주식 시장에서의 작업에 이어 현물 가격은 Brennan and Schwartz (1985)와 Paddock 외 (1988) 원유와 같은 지속적인 성장률을 지닌 기하학적 브라운 운동으로 모델링되었습니다. 평형 수준으로 다시 가격을 가지고하는 경향이있을 것이다 공급이나 수요 측면에서 그 가격 기반의 의사 결정을 관찰, 같은 딕 시트와 Pindyck (1994)와 같은 다른 저자는 프로세스가로 알려진에 의미 - 복귀 현물 가격을 모델링 선호 한 및 일정한 평균값. Ådland (2003)는 선도 시장의 유동성 부재로 현물 가격 모형의 사용을 주장하면서 운임에 대한 평균 반전 현물 가격 모형을 개발했다. 현물 가격의 이러한 단일 요소 모델은 가격의 행동에 대한 좋은 직관을 제공하지만 중요한 효과, 특히 contango에서 backwardation으로의 forward curves의 전환 및 만기와 관련하여 선물 계약의 변동성 감소를 포착하지 못합니다. Longstaff, Santa-Clara and Schwartz (1999)는 swaptions 시장에서 몇 가지 요인을 설명하지 않으면 차선책을 초래할 수 있음을 자세히 설명합니다. 이것을 설명하기 위해 Gibson과 Schwartz (1990)는 평균 - 반전 확률 적 편리율을 도입했다. 따라서 그들의 모델에는 전방 곡선에 충격을주는 두 가지 요인, 즉 현물 가격, 수준에 영향을주는 편의 수익률, 경사에 영향을 미치는 요인이 있습니다. 이 두 요인 모형은 Baker, Mayfield and Parsons (1998) 및 Schwartz and Smith (2000)와 같이 장기 및 단기 충격 측면에서 재 해석 될 수있다. 이 모델에서 현물 가격은 평균 반전 단기 계수와 장기적인 장기 요인에 충격을받습니다. 이 모델들은 모두 현물 가격 모델입니다. 그들은 관측 가능하고 가장 유동적 인 가격 인 현물 가격의 거동을 설명하려고합니다. 그런 다음 그들은 시장 가격의 위험 및 차익 거래없는 가격 책정을 도입하여이 과정에서 선물을 가격을 책정하고 전방 곡선에 대한 프로세스를 도출합니다. 역방향 접근법은 전체 진행 곡선을 기본 프로세스로 사용하는 것입니다. Miltersen과 슈워츠 (1998), Clewlow 및 스트릭 랜드 (2000)와 Sclavounos 및 Ellefsen (2009)는 다중 요소 히스, Jarrow와 모튼 (1992) 금리의 기간 구조에 대한 모델에서 영감을 같은 모델을 개발한다. 이는 순방향 곡선의 공분산 행렬을 소수의 직각 주성분으로 분해하는 것으로 구성됩니다. 현물 가격 프로세스는 순방향 곡선의 정면 가격으로 도출됩니다. 우리가 채택 할 접근 방식이지만, 주요 구성 요소 모양에 대한 파라 메트릭 가설을 만들고이를 공분산 행렬에 맞게 조정합니다. 3. 탐색 적 데이터 분석 상품 시장의 주요 특징에 대한 아이디어를 얻기 위해 서로 다른 상품의 과거 가격에 대한 분석을 시작합니다. Spot price 그림 1에서 우리는 최근 기간 동안 다른 상품의 현물 가격을 제시합니다. 원유와 같은 많은 시장에서이 현물 가격은 물적 거래가있는 전월 선물 계약의 가격으로 이해됩니다. 선적과 같은 다른 시장에서 현물 가격은 금융 선물 계약이 정산되는 다른 브로커의 현물 고정을 사용하여 매일 집계되는 지수입니다. 350 원유 난방유 천연 가스 TD3 해운 300 250 200 150 100 50 0 2005 2006 2007 2008 2009 2010 그림 1. 2005 년 1 월부터 2009 년 1 월까지 원유, 난방유, 천연 가스 및 TD3 운송 경로의 일별 현물 가격 지수 = 100 Forward curves 원유 천연 가스 S pot Forward price 160 16 S pot Forward price 14 140 12 가격 ($ / bbl) 가격 ($ / bbl) 120 100 10 8 80 6 60 40 2005 4 2006 2007 년 2008 년 2009 년 2 월 2005 년 2006 년 2007 년 2008 년 2009 년 2010 년 그림 2. 다른 날짜의 원유 및 천연 가스 매장량 및 가격 그림 2는 원유 및 천연 가스의 다른 날짜에 대한 전방 곡선을 나타냅니다. 우리는 전방 곡선의 수준이 현물 가격과 함께 이동하고 곡선이 contango와 backwardation 사이에서 전이하는 것을 관찰합니다. 게다가, 천연 가스의 전방 곡선은 계절적인 패턴을 가지고있다. 변동성 용어 구조 120 원유 난방유 천연 가스 TD3 해운 110 100 변동성 (%) 90 80 70 60 50 40 30 20 0 5 10 만기까지의 시간 τ (개월) 15 그림 3. 원유에 대한 선물 계약의 변동성 기간 구조 석유, 난방유, 천연 가스 및 TD3 해운 그림 3은 원유, 난방유, 천연 가스 및 TD3 해운에 대한 변동성의 용어 구조를 나타냅니다. 이는 만기까지 고정 된 계약의 역사적 변동성입니다. 우리는이 모든 상품들에 대해 단기 계약의 변동성이 사무엘 슨 (1965)의 가설과 일관되게 커브에서 벗어난 계약의 변동성보다 높다는 것을 관찰했다. 이것은 상품 시장의 중요한 특징이며 단기적으로는 장기적으로보다 비 탄력적이기 때문에 발생합니다. 유조선 시장이 포화 된 경우 한 달 안에 새로운 선박을 추가하는 것은 불가능하지만 다음 해에 증가하는 수요를 수용하기 위해 새로운 선박을 건조 할 수 있습니다. 상관 관계 구조 주어진 상품의 전방 가격은 독립적으로 움직이지 않습니다. 서로 다른 성숙도를 가진 계약의 상관 행렬을 관찰하면 이러한 움직임 간의 관계를 정량화 할 수 있습니다. 그림 4는 원유 계약에 대한 상관 행렬을 보여줍니다. 상관 관계 매트릭스는 서로 다른 계약간에 강한 상관 관계를 보여 주며, 앞선 월 계약과 60 개월 계약 간의 상관 관계는 84 %입니다. 그러나 전월 계약과 기타 계약 간의 상관 관계는 60 개월 계약과 이웃 계약 간의 상관 관계보다 더 빠르게 감소합니다. 1 상관 0.95 0.9 0.85 0.8 60 40 20 τ2 (개월) 원유 미래 주성분을 분석 0 0 20 10 30 40 50 60 τ1 (개월)도 4 상관 구조의 공동 운동의 구조에 대해 더욱 자세한 정보를 얻을 수 앞으로 가격 우리는 가격 시리즈의 주성분 분석 (PCA)을 수행 할 수 있습니다. 이것은 공분산 행렬의 고유 값과 고유 벡터를 찾는 것입니다. 고유치는 각 요소의 변동성 및 고유 벡터가 주요 구성 요소가 앞으로 곡선에 충격을주는 가중치로 해석 될 수 있습니다. 그림 5에서 원유 시장의 PCA 결과를 제시합니다. 이 결과에서 알 수 있듯이 지배적 인 요소가 첫 번째 요소인데, 이는 분산의 96.9 %를 차지합니다. 이 요소는 동일한 방향으로 가격을 전가하는 병렬 시프트 팩터입니다. 두 번째 요인은 분산의 2.8 %를 설명하며, 전진 커브의 프론트 엔드와 장단을 다른 신호로 충격을가함으로써 포워드 커브의 기울기에 영향을 미칩니다. 이것은 contango에서 backwardation으로의 전환을 설명합니다. 세 번째 요인은 앞쪽과 뒤쪽 끝을 적극적으로 충격을 가하고 곡선의 중간을 부정적으로 충격으로써 전진 커브의 볼록도에 영향을 미칩니다. 15 2 60 1 PC 무게 (τ) 1.5 40 휘발성 (%) 50 PC 1 PC 2 PC 3 30 0.5 20 0 10 -0.5 0 -1 1 2 3 주성분 4 5 처음 5 가지 주성분의 변동성 0 10 20 30 40 만기까지의 시간 τ (개월) 50 60 주요 구성 요소 가중치 그림 5. 원유 시장의 첫 번째 주요 구성 요소에 대한 변동성 및 가중치 4. 상품 선물의 2 요소 모델 날짜 t에서 현물 가격 S (t)를 정산하는 전방 곡선 F (t, T)를 관찰한다. S (t)는 시간 t에서 거래 가능한 상품의 현물 가격을 나타낼 수있다. 특정 위치) 또는 일일 공표 된 인덱스 값. 긴 전방 포지션이 날짜 t에서 입력되면, 그것은 T에서의 차이 S (T) - F (t, T)를받을 것이다. 재정 중재가 없으면 위험 중립적 인 측정 하에서 (Musiela and Rutkowski, 2004) T (T, T)) = Et * [S (T)] (2.1) 여기서 B (t, T ) 시간 T에서 성숙한 제로 쿠폰 채권의 시간 t 가격. 시간 t에서의 S의 전매 가격은 위험 중립 측정 하에서 시간 t에서 현물 가격의 기대치이며 시간 t에서의 정보가 주어진다. In some markets where the spot is storable, such as equities or currencies, there is a tight arbitrage enforcing the relationship between spot and forward prices. In markets where storage is limited, such as crude oil or shipping, the forward price is determined by supply and demand. It is not our goal here to impose a parametric model for the shape of the initial forward curve, which we take as given, but to give a model of its future stochastic evolution. Following Baker, Mayfield and Parsons (1998) and Schwartz and Smith (2000), we suggest a two-factor model for the stochastic evolution of the forward curve. We present this model as a forward curve model rather than a spot price model, considering that the commodity derivative markets are generally more liquid than their physical counterparts, and contain more information to calibrate on than the spot price. 16 We suggest the following two-factor model for the forward curve under the risk-neutral measure: dF (t , T ) = σ S e−α (T −t ) dWS (t ) + σ L dWL (t ) F (t , T ) dWS dWL = ρ dt (2.2) This is a four-parameter model and as we will show the parameters can be interpreted as follows: • σ S is the volatility of short-term shocks to the forward curve, • σ L is the volatility of long-term shocks, • α is the mean-reversion speed, quantifying how fast short-term shocks dissipate, • ρ is the correlation between short-term and long-term shocks. Covariance and correlation This model implies a covariance matrix between contracts that can be calculated as a function of the parameters • Covariance matrix: Σt (T1 , T2 ) = dF (t , T1 ) dF (t , T2 ) 1 Cov , dt F (t , T1 ) F (t , T2 ) 2 = (σ S e −α (T1 −t ) + ρσ L )(σ S e −α (T2 −t ) + ρσ L ) + (1 − ρ 2 )σ L (2.3) = Σ(τ 1,τ 2 ) where τ k = Tk − t is the time to maturity of the contract • Futures instantaneous volatility function: σ inst (t , T ) = (σ S e−α (T −t ) + ρσ L ) 2 + (1 − ρ 2 )σ L 2 = σ inst (τ ) • (2.4) 2 σ 0 = (σ S + ρσ L ) 2 + (1 − ρ 2 )σ L (2.5) Spot volatility: 17 • Correlation matrix: dF (t , T1 ) dF (t , T2 ) Σ(T1 , T2 ) , = F (t , T1 ) F (t , T2 ) σ inst (t , T1 )σ inst (t , T2 ) ρt (T1 , T2 ) = Corr (σ S e −α (T1 −t ) + ρσ L )(σ S e−α (T2 −t ) + ρσ L ) + (1 − ρ 2 )σ L 2 = 1/2 2 (σ S e −α (T1 −t ) + ρσ L )2 + (1 − ρ 2 )σ L = ρ (τ 1,τ 2 ) 1/2 2 (σ S e −α (T2 −t ) + ρσ L ) 2 + (1 − ρ 2 )σ L (2.6) All these quantities depend only on the time-to-maturities τ = T − t of the contracts involved, and not on time t. Implied spot price process In Appendix 2 we show that the spot price model consistent with this forward curve model is: dlog S (t ) = α ( µ (t ) − log S (t ))dt + σ S dW1 (t ) + σ L dW2 (t ) d µ (t ) = m(t )dt + σ L dW2 (t ) (2.7) i. e. the spot price is mean-reverting to a stochastic mean. This is equivalent to the Schwartz and Smith (2000) model which can be rewritten as µ dlog St = κ ξ + ξt − log St dt + σ χ dz χ + σ ξ dzξ κ dξ t = µξ dt + σ ξ dzξ (2.8) From equation (2.7) we can see that α can be interpreted as the speed of mean-reversion and σ L as the volatility of the long-term shocks. 5. Principal components analysis As discussed in Sclavounos and Ellefsen (2009), futures markets can be analyzed and modeled in a nonparametric way through principal components analysis (PCA) of the covariance matrix, leading to a multifactor Heath-Jarrow-Morton model of the form dF (t , T ) d = ∑ σ k (t , T )dWk , F (t , T ) k =1 dWk dWl = δ kl dt (2.9) Given the parametric model presented here, we can perform a PCA of the model’s covariance matrix and deduce the shape of its principal components. This will allow us to reformulate the model in terms of independent factors that can be interpreted in terms of their actions on the forward curve. 18 In the continuous setting we perform the Karhunen-Loève decomposition of the process following Basilevsky (1994). Let f (t, τ ) = F (t , t + τ ) be the constant-maturity forward with time-to-maturity τ . We want to decompose its evolution into: ∞ df (t, τ ) = µ (t, τ )dt + ∑ λk uk (τ )dzk f (t, τ ) k =1 (2.10) Where: • The zk are independent Brownian motions • The functions uk are the eigenvectors of the covariance matrix Σ(τ 1,τ 2 ) with associated eigenvalues λk : for some arbitrary maximal tenor τ max , τ max ∫ Σ(τ 1,τ 2 )uk (τ 2 )dτ 2 = λk uk (τ 1 ) 0 (2.11) τ max ∫u 2 k (τ )dτ = 1 0 We solve this eigenvector problem analytically in Appendix 3, and show that there are only two distinct functions uk (because it is a two-factor model), and they can be written in the form uk (τ ) = Ak e−ατ + Bk (2.12) where ( Ak , Bk ) and λk are solutions of the two-dimensional eigenvalue problem τ max 2 −ατ −ατ ∫ (σ S e 2 + ρσ Sσ L )e 2 dτ 2 A λk k = τ 0 Bk max 2 ( ρσ Sσ L e−ατ 2 + σ L )e −ατ 2 dτ 2 ∫ 0 τ max 2 (σ S e−ατ 2 + ρσ Sσ L )dτ 2 0 Ak τ max B −ατ 2 2 k ∫ ( ρσ Sσ L e + σ L )dτ 2 0 ∫ (2.13) The volatility of factor k is then related to λk by σ k = λk . The shape of the eigenfunctions is given in Figure 6 in the case of crude oil futures. We can notice that u1 corresponds to parallel shifts of the forward curve, whereas u2 corresponds to tilts. This is consistent with the two first factors observed doing a PCA of the historical covariance matrix (Sclavounos and Ellefsen, 2009). 19 1.5 u1(τ) u2(τ) 1 0.5 0 -0.5 0 10 20 30 TTM (m) 40 50 60 Figure 6. Shape of the eigenfunctions u1 (τ ) and u2 (τ ) for σ S = 18.1% , σ L = 23.3% , α = 0.842 yr −1 , ρ = 0.195 and τ max = 5 years This allows us to reformulate the evolution of the individual forward contract expiring at date T, in the riskneutral measure: dF (t , T ) = σ 1u1 (T − t )dz1 (t ) + σ 2u2 (T − t )dz2 (t ) F (t , T ) (2.14) Constant-maturity forward curve In Appendix 4 we show how this translates to the evolution of constant-maturity forward curve. We show that the constant-maturity futures price f (t, τ ) = F (t , t + ) can be written as: log f (t, τ ) = log F (0, t + τ ) + ψ 1 (t, τ ) + ψ 2 (t, τ ) + g1 (t ) + g 2 (t ) + u1 (τ ) f1 (t ) + u2 (τ ) f 2 (t ) (2.15) where: df k (t ) = −α k f k (t )dt + σ k dWk (t ) dg k (t ) = Bkα k f k (t )dt 1 dψ k (t, τ ) = − σ k2uk2 (t + τ )dt 2 Thereby we have decomposed the forward curve’s shape at time t into • Its initial shape F (0, t + τ ) , which under the risk-neutral measure is also its expected shape • A deterministic risk-neutral drift ψ 1 (t, τ ) +ψ 2 (t, τ ) ensuring that Et*[ F (t , T )] = F (0, T ) 20 (2.16) • A stochastic drift g1 (t ) + g 2 (t ) , independent of the maturity τ • Two independent mean-reverting factors f1 (t ) and f 2 (t ) (volatilities σ k and mean-reversion speeds α k ), giving rise to a parallel shift and a tilt, according to the shape of the factor weights uk (τ ) The spot price process S (t ) is given by the zero time-t o-maturity price f (t , 0) . This allows us to express the evolution of the forward curve as the result of shocks from two independent mean-reverting factor values. f1 , the parallel shift factor, affects the average level of the forward curve and is the dominant factor. f 2 , the tilt factor, affects the slope of the forward curve, as seen in Figure 7. u2(τ)f2 u1(τ)f1 u2(τ)f2 Figure 7. Effect of a positive parallel shift (left) and tilt (right) on the forward curve 6. Forward curve seasonality A number of commodities have seasonal prices. It appears because demand or supply is seasonal, and inventories are not sufficient to smooth this seasonality out over the year. Examples of seasonal commodities are heating oil and natural gas (winter heating demand), gasoline (summer driving season and different volatility requirements during summer and winter) and agricultural products (seasonal supply). This seasonality in spot prices is reflected in the forward prices because of market expectations. The difficulty when analyzing such forward prices is that the seasonality masks the underlying shifts in level and tilt that we are interested in. When considering such a seasonal commodity, the forward curve can be decomposed into a trend component and a seasonal component: log F (t , T ) = log FT (t , T ) + log FS (t , T ) 21 (2.17) The trend component FT (t , T ) represents the underlying non-seasonal forward curve, whereas the seasonal component FS (t , T ) , which for a given t is 1-year-periodic in T , represents the seasonal aspects of the curve. Pilipovic (2007) suggests a functional form that we have successfully applied to the natural gas and heating oil markets. In Section 2.10 we will show that this form is also consistent with the static shape of our three-factor model. For the trend component, log FT (t , T ) = ( A1e−α1 (T −t ) + B1 ) f1 + ( A2e −α 2 (T −t ) + B2 ) f 2 + ( A3e −2α3 (T −t ) + B3e−α3 (T −t ) + C3 ) f 3 (2.18) This functional form is flexible enough to reproduce the shapes of the underlying forward curve. For the seasonal component, we use sinusoidal seasonality with two harmonics (time must be measured in years) log FS (t , T ) = a1 cos(2π (T − t )) + b1 sin(2π (T − t )) + a2 cos(4π (T − t )) + b2 sin(4π (T − t )) (2.19) The only test of this model is how good the fit to the forward curve is. We find the parameters by leastsquares minimization for each day in the data set. A selection of forward curves is presented in Figure 8. While the fit is not perfect, the trend component seems to correctly capture the underlying trend, and that is what we are ultimately interested in. We then use this trend as the new forward curve, and carry out the rest of the calibration procedure on it. 7 9.5 Forward curve Fitted curve Trend component 6.5 Forward curve Fitted curve Trend component 9 8.5 6 8 5.5 7.5 5 4.5 7 0 10 20 30 40 50 6.5 60 May 21, 2004 0 10 20 30 May 15, 2008 Figure 8. Fitted Natural Gas forward curves on different dates 22 40 50 60 1 1 0.8 0.6 0.6 ρ ρ 0.8 0.4 0.4 0.2 0.2 60 0 60 40 40 20 τ2 (months) 0 0 10 30 20 40 50 60 20 τ2 (months) τ1 (months) 0 0 10 20 30 40 50 60 τ1 (months) After deseasonalizing Before deseasonalizing Figure 9. Correlation surfaces of Natural Gas futures before and after deseasonalizing the curves Figure 9 shows the effect of the procedure on the correlation surface. The results of calibrating the two-factor model to this new correlation surface are discussed below. 7. Market calibration To be successful the model needs to correctly reproduce the volatilities and instantaneous correlations of the traded instruments. We show that there is a good fit to the crude oil, tanker shipping, natural gas and heating oil markets. Method 1: Least squares fit of the covariance matrix In order to calibrate the model, we perform the following steps: 1. If the forward curve is seasonal (such as natural gas, gasoline, heating oil), deseasonalize it using the technique described above (Section 2.6), and keep only the non-seasonal part FT (t , T j ) 2. From the available set of contract prices F (t , T j ) , construct constant-maturity prices f (t, τ j ) by linear interpolation using log f (t, τ j ) = (t + τ j − T j ) log F (t , T j +1 ) + (T j +1 − t − τ j ) log F (t , T j ) T j +1 − T j , T j < t + τ j < T j +1 (2.20) 3. From observations of f (t, τ j ) at dates t1 . tM +1 , construct logarithmic returns net of roll yield and their mean value 23 f (ti, τ j ) ∂ log f (ti −1,τ j ) R(ti, τ j ) = log , − (ti − ti −1 ) f (ti −1,τ j ) ∂τ R (τ j ) = 1 M M ∑ R(t, τ i =1 i j ) (2.21) 4. Calculate the historical covariance matrix 1 ɶ Σ(τ j, τ k ) = M M ∑ ( R(t, τ i i =1 j ) − R (τ j ))( R(ti, τ k ) − R (τ k )) (2.22) 5. Find the parameters σ S , σ L , α , ρ that minimize the squared error: N ∑ Σσ αρ min σ S, σ L , , j , k =1 S, σ L, α , ρ ɶ (τ j, τ k ) − Σ(τ j, τ k ) 2 (2.23) The results are presented in Table 1 for the crude oil, shipping, natural gas and heating oil markets. The results indicate that a satisfactory fit to the volatility term structure and correlation surface can be obtained using the two-factor model presented here. The best calibration results are obtained for crude oil futures, which is arguably the most liquid market of the four. It is also interesting to note the differences between the values obtained. The short-term volatility of shipping futures is extremely high, at 143%, reflecting the high spot price volatility, but its long-term volatility is comparable to the other markets, at 28.7%. 24 Table 1. Calibration results for different commodity markets Crude oil (Nymex WTI) Tanker shipping (Imarex TD3) Contracts: Imarex TD3 futures Frequency: daily Frequency: weekly Source: Thomson Datastream Parameters Period: January 2005 – March 2009 Contracts: NYMEX WTI futures Data Period: April 2005 – October 2008 Source: Imarex σS σL α ρ σS σL α ρ 18.1% 23.3% 0.842 0.195 143% 28.7% 3.32 -0.01 120 33 Historical Model 32 Historical Model 31 100 Vol (% annual) Vol (% annual) Volatility 30 29 28 27 80 60 26 25 40 24 23 0 10 20 30 TTM (m) 40 50 20 60 Volatility term structure of crude oil futures Historical Model 1 0 2 4 6 TTM (m) 8 10 12 Volatility term structure of TD3 futures Historical Model 1 0.9 0.95 0.8 Correlation 0.9 0.7 0.6 0.85 0.5 0.4 15 0.8 60 10 40 20 τl (m) 0 0 10 20 30 40 50 5 60 τl (m) τk (m) 0 0 2 4 6 8 10 τk (m) Correlation surface of TD3 futures Correlation surface of crude oil futures 25 12 2.5 1.5 u1(τ) Principal Components u1(τ) 2 u2(τ) u2(τ) 1.5 1 1 0.5 0.5 0 0 -0.5 -1 -0.5 0 10 20 30 TTM (m) 40 50 60 -1.5 0 2 4 6 TTM (m) 8 10 12 Model principal components for crude oil Model principal components for TD3 futures Natural Gas (Nymex Henry Hub) Heating oil (Nymex New York Harbor) Period: May 2002 – August 2009 Contracts: NYMEX NG futures Contracts: NYMEX HO futures Frequency: daily Frequency: daily Source: Reuters Source: Reuters σS σL α ρ σS σL α ρ 53% Parameters Data Period: October 2002 – August 2009 17.3% 0.762 -0.172 27.6% 26.4% 1.386 0.228 65 40 Historical Model Historical Model 60 38 55 36 45 Vol (% annual) Vol (% annual) Volatility 50 40 35 30 34 32 30 25 28 20 15 0 10 20 30 TTM (m) 40 50 26 0 60 Volatility term structure of natural gas futures 2 4 6 8 10 TTM (m) 12 14 16 Volatility term structure of heating oil futures 26 18 1 Historical Model 1 Historical Model 0.98 0.8 0.96 0.94 0.6 Correlation 0.92 0.9 0.4 0.88 0.2 60 0.86 20 40 10 20 τl (m) 0 0 10 20 30 60 50 40 20 15 10 5 0 τk (m) τk (m) Correlati on surface of heating oil futures Correlation surface of natural gas futures 1 2 u1(τ) u1(τ) 0.8 Principal Components 0 τl (m) u2(τ) 0.6 u2(τ) 1.5 1 0.4 0.5 0.2 0 0 -0.5 -0.2 -1 -0.4 -0.6 0 10 20 30 TTM (m) 40 50 -1.5 0 60 Model principal components for natural gas 27 2 4 6 8 10 TTM (m) 12 14 16 18 Model principal components for heating oil Method 2: Calibration of the individual factors The first two principal components have a simple expression in this model, and can be used for calibration. The method is the same as above, but we replace steps 4 and 5 with: ɶ 4’. Calculate the PCA of the historical covariance matrix and extract the first two factor loadings u1 (τ j ) , ɶ u2 (τ j ) 5’. Calibrate the exponential functional form on each of the factors by least squares: N min Ak , Bk, α k ɶ ∑ u (τ j =1 k j ) − ( Ak e −α kτ j + Bk ) 2 (2.24) We present the results of this method for crude oil futures in Table 2 and Figure 10. Table 2. Principal component parameters for crude oil futures, using two calibration methods Principal Component 1 Principal Component 2 Method 1 Method 2 Method 1 Method 2 σ 54.91 % 54.91 % 9.53 % 9.50 % A 0.1218 0.1205 1.7639 1.7385 B 0.4177 0.4189 -0.4435 -0.5148 α 0.8422 0.8713 0.8422 0.6707 0.54 1.4 Historical Method 1 Method 2 0.52 Historical Method 1 Method 2 1.2 1 0.5 0.8 0.6 0.48 0.4 0.46 0.2 0 0.44 -0.2 0.42 -0.4 0.4 0 10 20 30 TTM (m) 40 Principal Component 1 50 60 -0.6 0 10 20 30 TTM (m) 40 50 60 Principal Component 2 Figure 10. Fit of the shape of the two principal components using the two different calibration methods 28 We see that the two methods give very close results, except that the second method allows for a differe nt value of α which gives a slightly better fit to the second principal component. It should be noted that Method 2 adds one extra free parameter by allowing α1 and α 2 to be different. 8. Forward risk premia – from the risk neutral to the objective measure The present model has been formulated under the risk-neutral measure. The prices evolve under the real measure. The change of measure from the risk-neutral to the real measure involves introducing a riskpremium λk for each of the Brownian motions Wk . We assume this risk premium to be constant. dWk → dWk + λk dt (2.25) This will affect the factor processes f k (t ) and g k (t ) studied in Section 2.5: t f k (t ) = σ k ∫ e−α k (t − s ) (dWk ( s ) + λk ds ) 0 t df k (t ) = σ k dWk (t ) + σ k λk dt + σ k ∫ (−α k e −α k ( t − s ) )(dWk ( s ) + λk ds ) (2.26) 0 df k (t ) = (σ k λk − α k f k )dt + σ k dWk (t ) = ( µ k − α k f k )dt + σ k dWk (t ) And for the drift process g k (t ) : t g k (t ) = Bkσ k ∫ (1 − e−α k ( t − s ) )(dWk ( s ) + λk ds ) 0 t dg k (t ) = Bkσ k dWk (t ) + λk dt − dWk (t ) + λk dt + ∫ (−α k )e −α k ( t − s ) (dWk ( s ) + λk ds )dt (2.27) 0 dg k (t ) = Bkα k f k (t )dt Hence the factor process f k (t ) follows an Ornstein-Uhlenbeck process mean-reverting to µk = σ k λk / α k instead of 0. The definition of g (t ) does not change. We let µk = σ k λk be the drift term for the factor k. The stochastic evolution of the forward price with tenor T can then be written as dF (t , T ) = ( µ1u1 (T − t ) + µ2u2 (T − t ))dt + σ 1u1 (T − t )dW1 + σ 2u2 (T − t )dW2 F (t , T ) (2.28) These results show how to incorporate drifts of the forward curve into the model. These can be based on historical evidence of drifts in prices or subjective evaluations of the expected future prices. This allows 29 valuation models of physical assets to take into account forecasts of future price evolution. Financial derivatives, however, will be valued under the risk-neutral measure. 9. Extension to three factors The model we have considered is sufficient to reproduce the volatility and correlation term structures of most forward markets. However, it only allows for certain movements of the forward curve, i. e. parallel shifts and tilts. As shown previously in the Principal Components Analysis, the forward curve does have other movements, and the third principal component is generally understood to correspond to changes in curvature. Certain strategies, such as a butterfly trade, are especially sensible to this kind of change. We suggest modeling the third principal component as u (τ ) = Ae−2ατ + Be−ατ + C (2.29) As is shown in Figure 11 it gives a good fit to the third principal component calculated from a historical covariance matrix. With parameters A and C positive and B negative the function u (τ ) will take positive values for small times-to-maturity, negative values for intermediate τ , and then positive values again, thereby affecting the convexity of the curve. Figure 11. Fit of the parametric third PC to the third PC from the covariance matrix (crude oil) 30 u3(τ)f3 u3(τ)f3 u3(τ)f3 Figure 12. Effect on the forward curve of a positive shock from the third principal component In order to study its interpretation we will consider its effect on the constant-maturity forward curve, as we did in Section 2.4 for the first two components. In Appendix 5 we show that the constant-maturity forward curve can be written as 2 log f (t, τ ) = log F (0, t + τ ) + ∑ (ψ k (t, τ ) + g k (t ) + uk (τ ) f k (t )) k =1 +ψ 3 (t, τ ) + B3e −α 3τ (2.30) g3 (t ) + C3 h3 (t ) + u3 (τ ) f 3 (t ) where df 3 (t ) = −2α 3 f 3 (t ) + σ 3 dW3 (t ) (Ornstein-Uhlenbeck process) dg 3 (t ) = α 3 ( f 3 (t ) − g3 (t ))dt g 3 (t ) = α 3 ∫ e−α3 ( t − s ) f3 ( s )ds t (2.31) 0 dh3 (t ) = 2α 3 f (t )dt t h3 (t ) = 2α 3 ∫ f 3 ( s )ds 0 The process f 3 (t ) is an Ornstein-Uhlenbeck process mean-reverting to zero with mean-reversion speed 2α 3 and volatility σ 3 . The processes g3 (t ) and h3 (t ) are stochastic drifts – integrals of f 3 (t ) with different weights. 10. Model of the static forward curve While the starting point of our modeling is that the initial forward curve F (0, T ) is given, there are situations where one would want to model this curve with a small number of parameters. Using the factor 31 model presented in this part, we can express the possible shapes of the curve when starting from an initial forward curve: N log f (t, τ ) = log F (0, t + τ ) + ∑ψ k (t, τ ) + g k (t ) + uk (τ ) f k (t ) (2.32) k =1 If we assume the initial forward curve to be flat, F (0,τ ) = F , this formulation simplifies to: N log f (t, τ ) = A(t ) + ∑ψ k (t, τ ) + uk (τ ) f k (t ) (2.33) k =1 Where A(t ) is a time-dependent scalar not depending on time-to-maturity τ and t 1 ψ k (t, τ ) = − ∫ σ k2 ( s, t + τ )ds 20 (2.34) This gives the possible shapes that can be taken by the forward curve given an initially flat curve. We can further simplify this by remarking that the first factor, the parallel shift factor, has a function u1 (τ ) that is almost constant, such that the constant term can be merged into the first factor value. Thereby the forward curve can be written as N f (t, τ ) = exp ∑ψ k (t, τ ) + uk (τ ) f k (t ) k =1 (2.35) Hence it can be described by N + 1 state variables: t , f1 . f N . Their initial values can be calibrated on the initial forward curve by calculating f k (0) = τ max ∫ u (τ ) log ( F (0,τ ) )dτ (2.36) k 0 This formulation also allows us to relate the average level of the forward curve and the first factor value by forming the geometric average weighted by u1 (τ ) : τ max u (τ ) F (t ) = exp ∫ w(τ ) log f (t, τ )dτ , w(τ ) = τ max 1 0 u (τ )dτ ∫ 0 32 1 (2.37) such that N τ max f1 (τ ) + ∑ ∫ u1 (τ )ψ k (t, τ )dτ k =1 0 F (t ) = exp τ max ∫ u1 (τ )dτ 0 (2.38) We can also examine the initial slope of the curve, that we will use to determine if the curve is in backwardation or contango: ∂f ∂τ = f (t , 0) τ =0 ∂ log f ∂τ (2.39) τ =0 and N ∂ψ k ∂uk ∂ log f =∑ + f k (t ) ∂τ ∂τ k =1 ∂τ (2.40) If we are considering a two-factor model, the first factor is almost flat such that its derivative is zero. In that case the only contribution comes from the second factor: ∂ log f ∂ψ 2 ∂u2 = + f 2 (t ) ∂τ ∂τ ∂τ (2.41) such that the initial slope of the forward curve is ∂f ∂τ τ =0 ∂ψ = f (t , 0) 2 ∂τ + τ =0 ∂u2 ∂τ τ =0 f 2 (t ) (2.42) Thus the value of f 2 (t ) determines the slope of the forward curve. 11. Applications of the market model Derivatives pricing The main application of stochastic models of forward curves is in derivatives pricing. The stochastic model that we have derived and calibrated allows for simple pricing of paper derivatives depending on the volatility of prices, such as European or Asian options written on the forward or spot price. In Part 3 we will derive analytical prices of commodity Asian options using the two-factor model derived here. 33 Real asset valuation and operation There are a number of physical assets whose value depends on commodity prices and forward curves. Oil or gas reservoirs are a simple example, but more complex assets such as refineries, power plants or oil in transit depend on these prices in a more complex way. Their value depends not only on the spot price but on the complete forward curve, and operational decisions should be made taking into account the possible future evolutions of the complete curve. The value of such an asset can be written as V (t , f (τ )) where f (τ ) is the current forward curve. If a twofactor model such as the one in this thesis is adopted, f (τ ) is a function of the initial forward curve F0 (τ ) , time t and the factor values f1 and f 2 , such that the value can be written V (t , f (τ )) = V (t , f1 , f 2 ) (2.43) The stochastic evolution of this value function can then be derived, using Ito’s formula and the independence of the factors, as dV = ∂V ∂V ∂V 1 ∂ 2V 2 1 ∂ 2V 2 dt + df1 + df 2 + df1 + df 2 ∂t ∂f1 ∂f 2 2 ∂f12 2 ∂f 22 ∂V ∂V ∂V 1 2 ∂ 2V 1 2 ∂ 2V ∂V ∂V = + ( µ1 − α1 f1 ) + ( µ2 − α 2 f 2 ) + σ1 + σ 2 2 dt + σ 1dW1 + σ 2 dW2 2 ∂f1 ∂f 2 2 ∂f1 2 ∂f 2 ∂f1 ∂f 2 ∂t (2.44) Associated with appropriate boundary conditions this allows for the calculation of the value of the real asset and the hedging of its value using the factors. In Part 4 we present the results of this methodology for a physical crude oil trade involving the shipment and possibly storage of crude oil. Risk evaluation Once a portfolio of paper and real assets has been valued, the risk of the portfolio can be evaluated using the market model presented here. We assume that given a forward curve f (τ ) and a date t the portfolio has a value V (t , f (τ )) . If we assume a two-factor model this value can be re-written as V (t , f1 , f 2 ) and its stochastic evolution as dV = µ (t , f1 , f 2 )dt + ∂V ∂V σ 1dW1 + σ 2 dW2 ∂f1 ∂f 2 (2.45) ∂V ∂V σ 1dW1 ( s ) + ∫ σ 2 dW2 ( s ) ∂f1 ∂f 2 0 0 (2.46) Thereby t t V (t ) = ∫ µ ( s, f1 ( s ), f 2 ( s ))ds + 0 Hence the expected value of V at a horizon t is 34 t t E [V (t ) ] = ∫ µ ( s, f1 ( s ), f 2 ( s ))ds (2.47) 0 and its standard deviation 1/2 2 t t ∂V 2 ∂V 2 2 Std[V (t )] = E ∫ σ 1 ds + ∫ σ 2 ds 0 ∂f1 ∂f 2 0 (2.48) These values can be calculated if the value of V as a function of the factors and time is known explicitly. Alternatively Monte Carlo simulation can be used, using the two independent processes f1 and f 2 , to estimate the complete distribution of V at the horizon time t. This Monte Carlo simulation will only require the simulation of two independent stochastic variables and not of each forward price separately. This information about the distribution of the portfolio value can be used to evaluate the risk of the position and calculate risk measures such as value-at-risk. Hedging As seen above a portfolio that depends on forward prices has, according to the two-factor model, a stochastic evolution that can be written dV = µ (t , f1 , f 2 )dt + ∂V ∂V σ 1dW1 + σ 2 dW2 ∂f1 ∂f 2 (2.49) = µ (t , f1 , f 2 )dt + δ1σ 1dW1 + δ 2σ 2 dW2 In order to hedge the risk related to factor k the portfolio must be complemented with a position of −δ k in ɶ the factor k. In that case the hedged portfolio V has the stochastic evolution ɶ dV = ( µ (t , f1 , f 2 ) − δ k ( µ k − α k f k ))dt + ∑ δ jσ j dW j (2.50) j ≠k Such a position in a specific factor can only be established with the traded futures F (t , T j ) , j = 1. N . The future with tenor T has the instantaneous evolution dF (t , T ) = F (t , T ) (σ 1u1 (T − t )dW1 + σ 2u2 (T − t )dW2 ) (2.51) Consider a portfolio with w j contracts F (t , T j ) , such that N dWk ⋅ dV = ∑ w j F (t , T j )uk (T j − t ) σ k dt j =1 35 (2.52) For this portfolio to hedge the factor f k while being unaffected by the factors f l , l ≠ k , the following equations must be satisfied: N ∑ w F (t , T )u (T j =1 j j k j (2.53) N ∑ w F (t , T )u (T j =1 − t) = 1 j j l j − t ) = 0, l ≠ k If the number of factors is smaller than N there are several solutions to the equations. If there are only two factors, this can be accomplished using two distinct contracts F (t , T1 ) and F (t , T2 ) . To hedge factor 1: u2 (T2 − t ) 1 w1 = F (t , T ) D(t , T , T ) w1 F (t , T1 )u1 (T1 − t ) + w2 F (t , T2 )u1 (T2 − t ) = 1 1 12 ⇒ w1 F (t , T1 )u2 (T1 − t ) + w2 F (t , T2 )u2 (T2 − t ) = 0 u (T − t ) 1 w =− 2 1 2 F (t , T2 ) D(t , T1 , T2 ) (2.54) where D (t , T1 , T2 ) = u1 (T1 − t )u2 (T2 − t ) − u2 (T1 − t )u1 (T2 − t ) . The portfolio with w1 contracts expiring at T1 and w2 contracts expiring at T2 replicates the stochastic part of the factor f1 . Similarly, the portfolio replicating the stochastic part of f 2 with these contracts is w1 = − u1 (T2 − t ) 1 F (t , T1 ) D(t , T1 , T2 ) u (T − t ) 1 w2 = 1 1 F (t , T2 ) D(t , T1 , T2 ) (2.55) It should be noted that using only two contracts makes the hedge very sensitive to these two contracts. If a continuous forward curve F (t , T ) is available, a hedge of f k can be formed using all the contracts if the following conditions are satisfied: τ max ∫ w(τ ) F (t , t + τ )uk (τ )dτ = 1 0 τ max ∫ (2.56) w(τ ) F (t , t + τ )ul (τ )dτ = 0 l ≠ k 0 A solution to this equation is then: w(τ ) = uk (τ ) F (t , t + τ ) 36 (2.57) 3. ASIAN OPTIONS ON COMMODITIES 1. Definitions and markets Most liquid commodity futures traded on exchanges settle on a specific day. For example, Brent futures trading on the InterContinental Exchange settle on the ICE Brent index price on the day following the last trading day of the futures contract. Futures with physical delivery, such as NYMEX WTI futures, do not have cash settlement but the options trading on them settle on their value on a specific day. In the case of freight derivatives the spot indices, published daily by the Baltic Exchange, are not considered liquid enough to be used for derivatives settlement. Given that there are relatively few spot transactions on a particular day, a big market participant might be able to manipulate the market to his favor over a period of a couple of days. To avoid this, the forward contracts settle on the average spot price over a month. This structure can also be found in over-the-counter swaps in other markets, such as crude oil or metals. Given a set of settlement dates T1 . TN (generally the trading days of a given month), the settlement price of the average contract settling on these dates will be FA (TN ; T1 . TN ) = 1 N N ∑ S (T ) k =1 (3.1) k This settlement price is also used for settling Asian options written on the same commodity. For example, the payoff of an Asian call option with strike K settling on the spot fixings on the dates T1 . TN is 1 C (TN , K ; T1 . TN ) = max N N ∑ S (T ) − K , 0 k =1 k (3.2) Asian options are very common in commodities – indeed they first appeared through commodity-linked bonds (Carr et al, 2008). They are popular not only because they avoid the problems of market manipulation as detailed above, but also because they are less expensive than their European counterparts. The Asian options we will consider are arithmetic average options with European exercise. Given a set of fixing dates T1 . TN , the option will pay off at date TN the value 1 C (TN , K ; T1 . TN ) = max N N ∑ S (T ) − K , 0 k k =1 1 P(TN , K ; T1 . TN ) = max K − N 37 ∑ S (Tk ), 0 k =1 (for a call) (3.3) N (for a put) 2. Literature on Asian options The existing literature on Asian options focuses on Asian options written on stock or foreign exchange rates. In this case the main effect of the averaging is in reducing the standard deviation of the payoff function. However, the distribution of the average of log-normal variables is not log-normal, and this is the main obstacle to pricing Asian options using the standard Black-Scholes framework. To tackle this, several techniques have been developed. Monte Carlo simulation can be used, such as in Kemna and Vorst (1990), Haykov (1993) and Joy et al. (1996). A partial differential equation depending on the spot price and the observed average price can be derived and solved numerically: Dewynne and Wilmott (1995) and Rogers and Shi (1995). Geman and Yor (1992) derive a semi-analytical expression for a spot price following geometric Brownian motion. Turnbull and Wakeman (1991) and Levy (1992) derive approximate expressions by matching the moments of a log-normal distribution with the moments of the average price distribution. A closed form expression is derived in Geman and Yor (1992) for a spot price following geometric Brownian motion. Approximate expressions have been obtained by Turnbull and Wakeman (1991) and Levy (1992). Haug (2006) presents these and other approximations for Asian options on futures. Koekebakker, Ådland and Sødal (2007) find an approximate expression for the Asian options trading in shipping, assuming the spot price follows geometric Brownian motion. Koekebakker and Ollmar (2005) use a one-factor forward curve model with time-varying volatility and derive an approximate process for the shipping forward freight agreement. A major issue in using these formulas for commodity futures options is that they assume geometric Brownian motion for the spot price, which is not consistent with a multi-factor model with mean-reverting factors. They also ignore the existence of a forward curve which gives the risk-neutral expectations of the spot price. 3. Approximate formulas under the two-factor model For option pricing we will work in the risk-neutral measure. The two-factor model of the forward curve is, as formulated in Part 2, dF (t , T ) = σ S e−α (T −t ) dWS + σ L dWL , F (t , T ) 38 dWS dWL = ρ dt (3.4) Consider an Asian forward contract FA settling on the average of the daily contracts F (⋅, T1 ). F (⋅, TN ) . The average price contract satisfies: FA (t ) = N 1 N ∑ F (t , T ) k k =1 1 dFA (t ) = N (3.5) N ∑ dF (t , T ) k k =1 where σ k (τ ) = 0 when τ < 0 (the contract has already settled, so its price is fixed). Then N dFA (t ) = FA (t ) ∑σ k =1 −α (Tk − t ) F (t , Tk ) Se N ∑ F (t , T ) N dWS + ∑σ k k =1 k =1 N L F (t , Tk ) ∑ F (t , T ) k =1 dWL (3.6) k Assumption 1: we assume the forward curve to be flat through the settlement period of the Asian contract: F (t , Tk ) = FA (t ) (all the daily contracts have the same price) If we make Assumption 1 then the above equation simplifies to the lognormal evolution dFA (t ) 1 = FA (t ) N N ∑σ k =1 S e−α (Tk −t ) dWS + σ L dWL (3.7) That is, the factor volatilities of FA are the average of the factor volatilities of the individual contracts. Assumption 2: Let us assume that the fixing dates are equally distributed: Tk = TN − ( N − k )h . Denote by c = TN − T1 the contract length. Then for t < T1 (pre-settlement): dFA (t ) σ S e−α (TN −t ) = FA (t ) N = σ S e−α (TN −t ) N −1 ∑ eα k =0 dWS + σ L dWL eα Nh − 1 dWS + σ L dWL N (eα h − 1) α = σ S e−α (TN −t ) kh (3.8) N (TN −T1 ) N −1 −1 e dWS + σ L dWL α (TN −T1 )/( N −1) N (e − 1) Assumption 3: The number of fixing dates N is large enough that N (eα (TN −T1 )/( N −1) − 1) ≈ α c . Then dFA (t ) eα c − 1 ≈ σ S e −α (TN −t ) dWS + σ L dWL αc FA (t ) 39 (3.9) Dynamics inside the settlement period An essential issue of Asian contracts is what happens inside the settlement period. As the contract enters the settlement period, its constituent prices are progressively discovered, and the uncertainty on its price at expiration diminishes. For example, the day before expiration, the only uncertainty is on the last price’s evolution during one day, which only contributes a small part to the average contract price. Furthermore, as part of the contract is priced, our Assumption 1 of a flat term structure F (t , Tk ) = FA (t ) becomes wrong. Indeed on date t , TM ≤ t < TM +1 , the spot prices S (T1 ). S (TM ) have been observed and they will not be equal to FA (t ) . We will therefore consider the observed average A(t ) ' and the adjusted average contract price FA (t ) : A(t ) = 1 M M ∑ S (t ), Tm ≤ t < Tm +1 k k =1 M 1 F (t ) = FA (t ) − A(t ) = N N ' A If we assume a flat term structure ahead, F (t , Tk ) = ' dFA (t ) = 1 N = 1 N (3.10) N ∑ k = M +1 F (t , Tk ) N ' FA (t ) for k ≥ m + 1 then N −M N ∑ k = M +1 dF (t , Tk ) N N FA' (t )(σ S e−α (Tk −t ) dWS +σ L dWL ) N −M k = M +1 ∑ ' dFA (t ) 1 = ' FA (t ) N − M N ∑ k = M +1 (3.11) σ S e−α (T −t ) dWS + σ L dWL k ' The adjusted average contract price FA (t ) has an approximately lognormal evolution with volatilities equal to the average of the volatilities. This, however, is only valid for TM ≤ t < TM +1 and assuming a flat term structure. The exact evolution of FA (t ) from t = T1 to t = TN is complex, and an exact derivation would have to follow the lines of Geman and Yor (1993). We do however get a good idea of the result by assuming that the already settled prices S (Tk ) and the daily forward prices F (t , Tk ) are all equal, in which case the average contract follows the evolution: dFA (t ) 1 = FA (t ) N N 1 N ∑ σ1 (Tk − t ) dW1 + k =1 N ∑σ k =1 2 (Tk − t ) dW2 (3.12) By definition σ k (τ ) = 0 for τ < 0 . In that case the average contract still has a lognormal distribution inside the settlement period, but with volatility decaying to 0 as shown in Figure 13. 40 150 1 day 1 month 3 months Vol (%) 100 50 0 0 2 4 6 8 Time to maturity (m) 10 12 Figure 13. Instantaneous volatility of the Asian contract as a function of time to maturity, for different contract lengths We want to quantify this contract’s stochastic evolution until maturity at TN . Consider a date t such that TM +1 ≤ TL ≤ t < TL +1 . Then: ' dFA (t ) σ S e −α (TN −t ) = ' FA (t ) N −M ≈ σ Se −α (TN − t ) N ∑ eα k = L +1 ( N −k ) h dWS + N −L σ L dWL N −M T −t eα (TN −t ) − 1 dWS + N ' σ L dWL ' α cM cM (3.13) (Assumption 3) ' where cM = TN − TM is the residual contract period. Pricing Asian options As discussed previously, in some markets the Asian forward contract/swap is traded in the market, and also settles on the average of the spot: FA (TN ; T1 . TN ) = Therefore we can rewrite the payoff of the option as 41 1 N N ∑ S (T ) k =1 k (3.14) 1 max N N ∑ S (T ) − K , 0 = max ( F (T k =1 k A N , T1 . TN ) − K , 0 ) (3.15) This payoff is a European call option written on the forward contract FA , which simplifies the problem considerably. We will denote by C (t , FA (t , T1 . TN ), K ) (3.16) the price, at time t, of a call option settling on the average of the spot prices at times T1 . TN , with strike K. Similarly P(t , FA (t , T1 . TN ), K ) denotes the price at time t of a put. Put-Call parity: because the options are standard European options written on a swap, put-call parity holds in the form C (t , FA (t ; T1 . TN ), K ) − P(t , FA (t ; T1 . TN ), K ) = B(t , TN )( FA (t ; T1 . TN ) − K ) (3.17) where B(t , TN ) is the discount factor. We will henceforth focus the discussion on calls. Black’s formula under time-dependent volatility We will begin by establishing Black’s (1976) formula for futures options under deterministic time-dependent volatility. We will follow the derivation of Musiela and Rutkowski (2008). Consider a futures process F (t , T ) with time-varying volatility: dF (t , T ) = µ (t , T )dt + σ (t , T )dWt F (t , T ) (3.18) We consider the futures option settling at date T on F (T , T ) with strike K: C (T , F ) = ( F − K ) + (3.19) Consider a self-financing futures strategy φ = ( g ( Ft , t ), h( Ft , t )) . Since the replicating portfolio φ is assumed to be self-financing, the wealth process V (t , F , φ ) , which equals V (t , Ft , φ ) = h( Ft , t ) Bt = C (t , Ft ) (3.20) satisfies dV = g ( Ft , t )dFt + h( Ft , t )dBt = rV (t , Ft )dt + µ (t , T ) Ft g ( Ft , t )dt + σ (t , T ) Ft g ( Ft , t )dWt 42 (3.21) If we assume that the function C is sufficiently smooth, we find by Ito’s lemma that ∂C ∂C 1 2 ∂ 2C ∂C + µ (t , T ) Ft + σ (t , T ) Ft 2 dC (t , Ft ) = dt + σ Ft dWt 2 ∂F 2 ∂F ∂F ∂t (3.22) Equating the values of V and C we find that we must have g ( Ft , t ) = ∂C (t , Ft ) ∂F (3.23) And ∂C 1 2 ∂ 2C + σ (t , T ) F 2 − rC = 0 ∂t 2 ∂F 2 C (T , F ) = ( F − K ) + (3.24) The solution of this partial differential equation is C (t , F (t , T )) = B(t , T ) [ F (t , T ) N (d1 ) − KN (d 2 ) ] (3.25) 1 ln( F (t , T ) / K ) + (T − t )σ Black (t , T ) 2 2 d1 = , d 2 = d1 − σ Black (t , T ) T − t σ Black (t , T ) T − t (3.26) where T 1 2 σ B lack (t , T ) = ∫ σ ( s, T ) ds T −t t 2 (3.27) Hence the formula for a call written on a futures contract with time-varying volatility is the same as Black’s (1976) formula for a futures option, except that the volatility is replaced with the root-mean-square of the instantaneous volatilities during the period. Instantaneous volatilities Let us now calculate the instantaneous volatilities of FA (t ) pre - and in-settlement, based on the stochastic evolution derived in equations (3.9) and (3.13). For t < T1 2 αc −1 −α (TN −t ) e 2 σ A (t , TN ) = σ S e + ρσ L + (1 − ρ 2 )σ L αc and in-settlement, for T1 ≤ t ≤ TN : 43 (3.28) 2 2 1 − e −α (TN −t ) T −t 2 T −t σ (t , TN ) = σ S + ρ N ' σ L + (1 − ρ 2 ) N ' σ L ' α cM cM cM ' A (3.29) Black volatilities Given these instantaneous volatility functions, we can calculate the Black volatility, i. e. the standard deviation of the contract price at expiration: σ Black (t , T ) 2 = T 1 2 σ A ( s, T )ds T∫ t (3.30) In-settlement Consider the case when dates T1 . TM have been priced, such that we are in fact considering the adjusted contract ' FA (t ) = FA (t ) − 1 N M ∑ S (Tk ) = k =1 1 N N ∑ k = M +1 F (t , Tk ) (3.31) In Appendix 6 we show that its Black volatility is: Such that the square of the Black volatility is given by: σ Black (t , T ) 2 = = T 1 2 ∫ σ A (s, T )ds T −t t 2 σS 2 1 − e−α (T −t ) 1 1 − e −2α (T −t ) + 1 − +⋯ ' α 2 cM 2 α T − t 2α T −t 2 ρσ Sσ L T − t e −α (T −t ) 1 1 − e−α (T −t ) + −2 +⋯ ' α cM 2 2 α α T −t (3.32) 1 (T − t ) 2 2 σL ' 3 cM 2 In the case when α c ≪ 1 this simplifies to σ Black (t , T ) ≈ 1 2 2 σ S + 2 ρσ Sσ L + σ L 3 Pre-settlement In Appendix 6 we show that 44 (3.33) 2 σ Black (t , T ) = 1 112 T −T 2 2 σ A ( s, T )ds = ∫ ∫ σ A ( s, T )ds + T − t1 σ Black (T1 , T ) T −t t T −t t T T (3.34) where 2 eα c − 1 e −2α (T −T1 ) − e −2α (T −t ) eα c − 1 e −α (T −T1 ) − e−α (T −t ) 2 σ ( s, T )ds = σ + 2 ρσ Sσ L + (T1 − t )σ L (3.35) ∫ 2α αc α αc t T1 2 A 2 S 2 and σ Black (T1 , T ) is given by equation (3.32). If α c ≪ 1 , 2 σ Black (t , T ) 2 ≈ σ S c 1 − e−2α (T1 −t ) 1 c 1 − e −α (T1 −t ) 1 2c 2 + + 2 ρσ Sσ L + + σ L 1 − (3.36) αc T − t 2α c 3 T −t 3 3 T −t Option price Once the Black volatility is known, pricing futures options is simply a matter of applying the equation (3.26) to the process FA (t , T ) . If the option is pre-settlement, C (t , FA (t , T ), K ) = B (t , T ) [ FA (t , T ) N (d1 ) − KN (d 2 ) ] P(t , FA (t , T ), K ) = B(t , T ) [ KN (−d 2 ) − FA (t , T ) N (−d1 )] (3.37) where 1 log( FA (t , T ) / K ) + (T − t )σ Black (t , T )2 2 , d 2 = d1 − σ Black (t , T ) T − t d1 = σ Black (t , T ) T − t (3.38) If it is in-settlement, introduce the already priced average. A(t ) = 1 M M ∑ S (t ), k =1 k TM ≤ t < TM +1 (3.39) The payoff at expiration can be rewritten as ' ' M t − T1 max ( FA (T ) − K , 0 ) = max FA (T ) − K − A(t ) , 0 ≈ max FA (T ) − K − A(t ) , 0 (3.40) N c ' such that the option can be considered to be written on the log-normal process FA (t ) with an adjusted strike K − A(t ) M / N . If K − A(t ) M / N ≥ 0 . The prices of the options are therefore: 45 t − T1 t − T1 C (t , FA (t , T ), K ) = B(t , T ) FA (t , T ) − A(t ) N (d1 ) − K − A(t ) N (d 2 ) c c t − T1 t − T1 P(t , FA (t , T ), K ) = B(t , T ) K − A(t ) N (−d 2 ) − FA (t , T ) − A(t ) N (−d1 ) c c (3.41) where: t − T1 FA (t , T ) − c A(t ) 1 ' 2 log + (T − t )σ Black (t , T ) t −T 1 2 K− A(t ) c ' d1 = , d 2 = d1 − σ Black (t , T ) T − t ' σ Black (t , T ) T − t (3.42) ' and σ Black (t , T ) is given by equation (3.32). If K − A(t ) M / N < 0 , the average over the contract period will always be larger than the strike, such that the call option will always be exercised and the put option will never be exercised. Thus C (t , FA (t , T ), K ) = B(t , T )( FA (t , T ) − K ) P(t , FA (t , T ), K ) = 0 (3.43) 4. Comparison to other Asian option models and market prices We examine two other Asian option models that are commonly used in the freight options market: the Levy (1992) approximation and the Koekebakker, Adland and Sodal (2007) formula. These models both assume that the spot price follows geometric Brownian motion under the risk-neutral measure: dS (t ) = λ dt + σ dWt S (t ) (3.44) Let Tk = T1 + (k − 1)h be the settlement period dates and define the observed running average A(t ) as in equation (3.10) above. Also let M (t ) = A(TN ) − A(t ) m and N v(t ) = ln E * M (t ) 2 − 2 ln E * [ M (t ) ] 1 m ln E * M (t ) 2 − ln K − A(t ) 2 N d1 = , d 2 = d1 − v(t ) v(t ) 46 (3.45) Then the Levy (1992) approximation is, for Tm ≤ t < Tm +1 , m C ( S (t ), A(t ), t ) = e − rτ E * [ M (t ) ] N (d1 ) − K − A(t ) N (d 2 ) N P( S (t ), A(t ), t ) = e − rτ * m E [ M (t ) ][ N (d1 ) − 1] − K − A(t ) [ N (d 2 ) − 1] N (3.46) Expressions for E *[ M (t )] and E *[ M (t ) 2 ] are given in the appendix of the original article. Koekebakker, Ådland and Sødal (2007) arrive at an option price along the following lines: 1. Assume geometric Brownian motion for the spot under the risk-neutral measure 2. Derive what the average forward price should be, consistent with this spot process FA (t ; T1 . TN ) = 1 N N ∑ Et*[S (Ti )] = St i =1 t ≤ T1 St N λ (T −t ) ∑ei N i = M +1 F ( t, T1,TN ) (3.47) dFA (t ; T1 . TN ) = σ (t )dWt where FA (t ; T1 . TN ) 3. Derive the forward’s approximate evolution: σ σ (t ) = σ eλ (TN −t ) 1 − e− λ N ∆ N 1 − e − λ∆ σ t ≤ T1 ≈ N −M TM < t < TM +1 TM < t < TM +1 σ N (3.48) 4. Use Black (1976) with time-varying volatility to price the option The result is as follows: for t < T1 (pre-settlement) let σ F = σ (T1 − t ) + R( N )(TN − T1 ) R( N ) = 1− 3 1 + 2N 2N 2 3 3− N For t = TM (in settlement) 47 (3.49) (3.50) σ 1 M −1 M σF = (TN − TM ) 1 − N 1 − N + 2 N 3 1/2 (3.51) Then C (t , FA (t ; T1 . TN ), K ) = B(t , TN ) [ FA (t ; T1 . TN ) N (d1 ) − KN (d 2 )] P(t , FA (t ; T1 . TN ), K ) = B(t , TN ) [ KN (−d 2 ) − FA (t ; T1 . TN ) N (−d1 ) ] (3.52) where B(t , TN ) is the discount factor and F (t ; T1 . TN ) 1 2 ln A + σF K 2 d1 = , d 2 = d1 − σ F σF (3.53) We now evaluate option premia based on these models. The volatility inputs to the models are evaluated based on the historical volatilities estimated from time series of prices of shipping futures contracts. We compare these to the option premia quoted by Imarex in their weekly Imarex Freight Options reports for the TD3 route. The main parameters are given in Table 3.. Table 3. Parameters on December 8, 2008 Date December 8, 2008 3-month LIBOR 2.19% Spot price (WS) 91.73 Observed average spot 12/1 – 12/5 (WS) 75.88 The key input to each of the models is the volatility. We use • For Levy and KAS, the historical volatility of the spot, using weekly log returns over the year 2008. This is evaluated to 148.4%. • We calibrate the two-factor forward curve model to the estimated historical covariance matrix for the year 2008. 48 160 Historical Model 140 Parameters: σ S = 172.4% Vol (% annual) 120 σ L = 34.8% 100 α = 3.245 80 ρ = 0.21 60 40 20 0 2 4 6 TTM (m) 8 10 12 Figure 14. Two-factor model fitted to historical volatilities Table 4. Model Call option ATM prices on December 8, 2008 compared to market Contract FFA Imarex Premium ATM Model option premium ATM (WS) Year Period WS WS Levy KAS Two-factor 2008 Dec 81 5.6 4.91 5.45 6.76 2009 Jan 59 8.8 10.69 10.56 11.07 2009 Feb 56 9.2 13.60 13.51 12.87 2009 Mar 46 8.2 13.34 13.28 11.52 2009 Apr 45 8.2 14.98 14.93 11.93 2009 Q2 45 8.5 16.48 16.44 12.27 2009 Q3 44 9.3 19.82 19.79 12.65 2009 Q4 48 11 24.66 24.63 14.18 2010 CAL 78 16.6 48.26 48.23 24.10 The results are listed in Table 4. We note that the Levy and KAS prices are very similar. This is not surprising as they are based on the same model and parameters for the underlying, only with different ways of approximating the option premium. The option premia obtained with the term structure of volatility from Figure 14 are lower than the Levy and KAS premia. This is obvious from Figure 15: the difference between KAS and the present study is the 49 volatility input to the Black formula - σ F and σ Black respectively – which are the root-mean-square of the time-dependent instantaneous volatility σ (t ) . Since the historical volatility of futures with increasing tenors is lower than that of the spot, the call premium from the present study will be lower than the KAS premium. 180 Two-factor KAS & Levy 160 140 Vol (%) 120 100 80 60 40 20 0 0 2 4 6 8 Time to maturity (m) 10 12 Figure 15. Asian volatility term structure for the two-factor model and the KAS & Levy models Comparing these premia to market prices quoted by Imarex, the market prices are lower than the prices obtained from the present model supplied with the historical volatilities – namely the market implies a lower future volatility than the historical volatility over 2008. Yet, 2008 was a particularly volatile year in the freight market, giving a high historical volatility, whereas the market might be expecting 2009 to be calmer. Implied volatilities – calibration of the models to market prices Volatility is the most important parameter of an option pricing formula. Considering that the other parameters are observable, a quoted market price for an option implies a value for this parameter. It is therefore important to be able to back out this parameter from the options premia observed in the market, resulting in the implied volatility. This follows by solving the equation C (t , FA (t , T ), T , K , r , σ ) = Cmarket (t , T , K ) (3.54) for σ or possibly several parameters that enter the definition of σ . Levy (1992) The formula (3.46) has a complicated dependence in σ and a numerical technique must be used to back out the volatility of the spot from the option price. As this is a single parameter, to each option price C (T , K ) 50 there corresponds an implied volatility σ Levy (T , K ) . The volatility of the spot and the options implied volatility are not the same because the market doesn’t follow the assumptions of the model. In particular, there is a term structure of volatility which is not consistent with the geometric Brownian motion model for the spot price. Imarex quotes these implied volatilities in their Freight Options reports. Koekebakker, Adland and Sødal (2007) The formula (3.52) is just Black (1976) with a tweaked volatility input. Extracting implied volatilities from the Black formula is standard, giving rise to σ Black (T , K ) . The implied volatility of the spot is then σ KAS (T , K ) = σ Black (T , K ) T −t g (t , T1 ,… , TN ) (3.55) where, consistently with equations (3.49) and (3.51), T1 − t + R ( N )(TN − T1 ) t < T1 g (t , T1 ,… , TN ) = 1 1 M − 1 M 3 (TN − TM ) 1 − N 1 − N + 2 N T1 ≤ t = TM (3.56) As with the Levy model, different implied volatilities are obtained for each maturity and strike. Two-factor model The input to the Black formula is the time dependent volatility σ A (t , T ) modeled according to (3.28) and (3.29). The parameters in the model need to be estimated from the term structure of options prices. Rebonato (2002) discusses in depth the calibration of the LIBOR market model to traded options in the interest rate markets, and much of his discussion applies here. σ Black (T ) for at-the-money options is first obtained from options market prices, and the parameters of the model are estimated by nonlinear least squares: ∑ σ αρ min σ S, σ L , , Model Black (T , σ S , σ L , α , ρ ) − σ Black (T ) 2 (3.57) T where the summation is over all available liquid option maturities. It should be noted that unlike the Levy and KAS models which estimate one implied volatility per options contract, the present model estimates four parameters from all liquid options prices by a nonlinear least squares technique, which is more parsimonious, but can lead to inaccuracies if the model is not suitable. 51 Note on calibrating a multi-factor model to implied volatilities The two-factor model that we present in this article is intended to reproduce not only the term structure of volatilities but also the correlation surface between the contracts. When pricing vanilla options, however, only the term structure of volatilities matters and the correlations are irrelevant. Thus, when calibrating the four parameters of the model to implied volatilities alone one cannot expect to correctly reproduce the correlation structure. However, considering that we have observed historical estimates of the parameter ρ to be close to 0, a simple solution consists in fixing this parameter to 0 and calibrating σ S , σ L , α to the implied volatility term structure. As long as this produces a good fit, vanilla options will be priced correctly. Results Based on the prices published in the Imarex Freight Options report on Dec 8, 2008, we extract implied parameters for the different models. The results are presented in Table 5. We fit the two-factor model (2.2) to the market prices using the procedure described above, and the result of the optimization is displayed in Figure 16. We can see that the two-factor model gives a very good fit to the option market prices. Figure 16. Black implied volatilities from market quoted options and calibrated two-factor model. Market prices are from the Imarex Freight Options report on December 8, 2008 52 Table 5. Implied volatilities for the different models on December 8, 2008 Contract Imarex Premium Implied volatilities σ Levy (T ) σ KAS (T ) KAS Black σ Black (T ) 165% 169.20% 152.53% 67.96% σS 177.25% 8.8 125% 121.88% 123.32% 97.43% σL 47.68% Feb 9.2 100% 99.60% 100.23% 87.51% α 8.7 2009 Mar 8.2 90% 89.92% 90.31% 81.30% ρ 0 2009 Apr 8.2 80% 79.52% 79.76% 73.86% 2009 May 8.6 75% 74.98% 75.16% 70.79% 2009 Jun 8.7 70% 69.60% 69.74% 66.14% 2009 Jul 8.7 66% 65.98% 66.10% 63.03% 2009 Aug 9.3 66% 66.27% 66.38% 63.76% 2009 Sep 9.8 66% 66.05% 66.14% 63.81% 2009 Oct 10.8 63% 63.44% 63.52% 61.49% 2009 Nov 11.1 62% 62.46% 62.53% 60.70% 2009 Dec 11.5 62% 62.10% 62.16% 60.42% Year Month WS Imarex report 2008 Dec 5.6 2009 Jan 2009 Levy Two-factor 5. Hedging of Asian options Greeks of Asian options When writing options, the seller may be interested in delta-hedging his portfolio with the underlying to construct a delta-neutral position. The question is what position to take in the underlying to hedge the option: the delta. The hedge ratio changes with the price of the underlying, time and volatility and must be adjusted regularly, leading to dynamic hedging strategies. When hedging a number of questions must be addressed: what is the underlying? What instruments are we going to hedge with? The options are written on spot, but in shipping the spot is ton-miles that are not be possible to buy and hold, nor short, since it is a service. In shipping the forward contracts that are trading in the market, the Forward Freight Agreements, settle on the average of the spot, and should therefore be used as hedging instruments. 53 The Greeks follow upon differentiation of the Black formula. Let d1 and d 2 be defined as in Section 3.2, then we have for a call option: ∆= ∂C = e− r (T −t ) N (d1 ) ∂FA Γ= n(d1 ) ∂ 2C = e − r (T − t ) 2 ∂FA FAσ B T − t dC ∂C ∂σ B ∂C Fσ ∂σ = + = rC − e − r (T −t ) A B n(d1 ) + B Vega ∂t ∂t ∂σ B ∂t dt 2 T −t ∂C = FAe− r (T −t ) T − tn(d1 ) Vega = ∂σ B θ= ρ= (3.58) ∂C = −(T − t )C ∂r Because the volatility is time dependent, the theta θ of the option also depends on the temporal variation of the Black volatility: T 2 (T − t )σ B (t , T ) = ∫ σ 2 ( s, T )ds t (3.59) Differentiating this with respect to t and rearranging we get 2 ∂σ B (t , T ) σ B (t , T ) − σ 2 (t , T ) = ∂t 2(T − t )σ B (t , T ) (3.60) which can be calculated using the formulas in Section 3. Delta-hedging an option position When a forward contract with the same settlement period as the Asian option is available, that contract should be used to delta-hedge the option position to avoid basis r isk. The position to be taken in this contract is then given by the previous formula. The position to take in the contract FA (t , T ) to hedge a short call position settling on the same period is ∆ C = e − r (T −t ) N (d1 ) (3.61) In other markets such a contract is not available. In the crude oil market, for example, the futures contracts settle on a single date while Asian options will settle on the trading days within a month. Luckily, these 54 contracts are highly correlated and we can quantify the required number of contracts using the two-factor model. The instantaneous evolution of the call price is dC = ∂C ∂C dt + dFA ∂t ∂FA ( = θ dt + ∆ FA FA (t , T ) σ (t , T )dWS + σ (t , T )dWL A S A L ) (3.62) ' Where FA (t , T ) = FA (t ; T1 . TN ) pre-settlement and FA (t , T ) = FA (t ; TM +1 . TN ) in-settlement, and eα c − 1 σ S e −α (T −t ) t < T1 αc A σ S (t , T ) = α ( T −t ) −1 σ e −α (T −t ) e TM ≤ t < TM +1 S ' α cM t < T1 σL A σ L (t , T ) = T − t c ' σ L TM ≤ t < TM +1 M (3.63) A daily contract settling on the date T has the stochastic evolution dF (t , T ) = σ S e−α (T −t ) dWS + σ L dWL F (t , T ) (3.64) We can use two of these contracts to hedge the call. Assume we take a position w1 in contract F (t , T1 ) and w2 in contract F (t , T2 ) , then the hedge of a short call must satisfy: w1 F (t , T1 )σ S e −α (T1 −t ) + w2 F (t , T2 )σ S e−α (T2 −t ) = ∆ FA FA (t , T )σ SA (t , T ) A ( w1 F (t , T1 ) + w2 F (t , T2 ))σ L = ∆ FA FA (t , T )σ L (t , T ) (3.65) which yields A σ SA (t , T ) σ L (t , T ) −α (T −t ) − e σL FA (t , T ) σ S 2 w1 = ∆ FA F (t , T1 ) e −α (T1 −t ) − e −α (T2 −t ) A σ SA (t , T ) σ L (t , T ) −α (T −t ) − e σL FA (t , T ) σ S 1 w2 = ∆ FA F (t , T2 ) e −α (T2 −t ) − e−α (T1 −t ) 55 (3.66) 6. Dependence of the Asian option price on the parameters Carr, Ewald and Xiao (2008) establish that in the Black-Scholes framework, the premium of an arithmetic average Asian call option written on stock increases with volatility. They also show that this is not a trivial result and does not hold outside the Black-Scholes assumption, for example using a binomial model. In the model presented here there is not a single volatility parameter but four parameters governing the term structure of volatility. We will study the dependence of the option premium on these four parameters. This has an important impact on option risk management, given that the implied term structure of volatility can change stochastically over time, thereby affecting prices. The dependence of the option premium on the volatility parameters is through σ Black , therefore we can write ∂C ∂C ∂σ B = , ∂σ S ∂σ B ∂σ S ∂C ∂C ∂σ B = . ∂σ L ∂σ B ∂σ L (3.67) We calculate the sensitivity of the Black volatility on the four parameters, in the simplified case when α c ≪ 1 . In-settlement, for TM ≤ t < TM +1 , σ B (t , T ) ≈ ∂σ B σ S + ρσ L = , ∂σ S 3σ B 1 2 2 σ S + 2 ρσ Sσ L + σ L 3 ∂σ B σ L + ρσ S = , ∂σ L 3σ B ∂σ B = 0, ∂α (3.68) ∂σ B σ Sσ L = ∂ρ 3σ B (3.69) And pre-settlement, t < T1 , 2 σ B (t , T ) 2 ≈ σ S c 1 − e −2α (T1 −t ) 1 c 1 − e−α (T1 −t ) 1 2c 2 + + 2 ρσ Sσ L + + σ L 1 − (3.70) T − t 2α c T −t 3 3 αc 3 T −t ∂σ B σ S c 1 − e−2α (T1 −t ) 1 ρσ L c 1 − e−α (T1 −t ) 1 = + + + ∂σ S σ B T − t 2α c 3 σB T −t 3 αc ∂σ B ρσ S c 1 − e−α (T1 −t ) 1 σ L T1 − t 1 = + + + ∂σ L 3 σ B c 3 σ B T − t αc T1 − t 1 −2α (T1 −t ) T − t 1 ∂σ B σ S c = + e − 1 + ρσ L 1 + e−α (T1 −t ) − 1 σ S ∂α σ B 2α (T − t ) c 2 2 c ∂σ B σ Sσ L c 1 − e−α (T1 −t ) 1 = + ∂ρ 3 σ B T − t αc 56 (3.71) Let us examine the sign of these quantities. In-settlement, we have, on the condition that σ S , ρ , σ L ≥ 0 , all the derivatives are non-negative. For the pre-settlement values, ∂σ B c σS ρσ L = g (2α ) + g (α ) ∂σ S T − t σ B σB ( (3.72) ) where g (α ) = 1 − e −α (T1 −t ) / (α c ) + 1/ 3 is decreasing in α , such that σ + ρσ L ∂σ B c ≥ ≥0 g (2α ) S ∂σ S T − t σB (3.73) ρσ S + σ L ∂σ B c ρσ S c σ = g (α ) + L g (0) ≥ g (α ) ≥0 ∂σ S T − t σ B σB σB T −t (3.74) For σ L , For α we can see that σ Black (t , T )2 is decreasing with α . The Black volatility is increasing in ρ without any conditions on the volatilities. Hence we have proven that, if we assume σ S ≥ 0, ρ ≥ 0, σ L ≥ 0 , the Black volatility and the call option premium are increasing in the parameters σ S , σ L and ρ , and decreasing in α . 57 4. THE FLOATING STORAGE TRADE 1. Introduction After the collapse of oil and shipping prices in mid 2008, floating storage, i. e. storing crude oil or products in idle tankers, became a viable opportunity. This was made possible by a steep contango of the crude oil forward curve combined with low freight rates. This is only one example of what international oil trading consists of: if a price difference – in time or space – is higher than the shipping and capital cost involved, then there is an arbitrage opportunity that can be exploited. Two very basic examples are: • Shipping crude oil from Nigeria to the US • Storing crude oil in storage tanks in Cushing, OK when the WTI forward curve is in contango. Often several opportunities present themselves to an oil trader, and the volatility of the associated forward curves makes it possible that these opportunities could evolve during the voyage. For example, a ship leaving Nigeria with crude oil has the option of going either to the United States or to Europe, and the trader doesn’t necessarily have to make that choice immediately – he can choose to stay on a northward course in the midAtlantic and defer the choice of destination port to a later date, when one option will be significantly more interesting than the other. The ship can also choose to speed up or slow down to control its fuel consumption and arrive at an optimal date. When choosing to keep his options open, the oil trader is exposed to movements in the forward curves. The existence of liquid futures and options markets at several key locations makes it possible to hedge this exposure partly, thereby reducing risks. Our aim is to create a framework for analyzing such trading strategies and derive the optimal route that a ship should follow. This framework can then be used to evaluate expected return and risk beforehand, to find the optimal route that the ship should follow, and to derive hedge ratios to hedge the exposure to the dominant risk factors. We will concentrate on a simple problem: the cross Atlantic crude oil arbitrage with possibility of floating storage, and present the results for this. We will then proceed to generalize the framework to a general optimal trading problem. 2. The floating storage problem We consider a ship at location X at time t. The characteristics of the ship and the shipping route are given in Table 6. The problem we are considering is as follows: 1. At date t = 0 the decision is made to load the tanker with a cargo at the spot price S0 58 2. The cargo is loaded at date τload. 3. The ship then sails at the constant speed u to the destination port. 4. Upon arrival, and until the date when the cargo is actually delivered, the ship is anchored at the destination port X Xp At destination Waiting for cargo to load Sailing, speed u X0 τ load τ load + d(X0, X p ) u Buy cargo, price S0 t τ* : deliver cargo at delivery port The daily cost paid for the shipping is then, at date t, t ≤ τ load 0 g (t ) = H + B ⋅ FC (u ) τ load ≤ t ≤ τ load + d ( X 0 , X P ) / u H + B ⋅ FC t ≥ τ load + d ( X 0 , X P ) / u a (4.1) The cargo is not necessarily sold into the market immediately. At any location X and time t the decision can be made to sell the cargo for delivery τ days forward, at the price F (t, τ ) . Exercise profit We define the exercise profit Ω(t , F (τ )) as the profit that can be earned on the cargo if the ship is at location X(t) and the forward curve is F (τ ) , by committing to a specific delivery price sometime in the future and sailing to deliver at that time. In effect, the trader gives up the possibility of changing delivery time. At exercise, one chooses a time-to-delivery τ . For one choice of this parameter the profit is τ sail = d ( X (t ), X P ) / u ω (τ ) = F (τ ) − S0 − Hτ − B ⋅ ( FC (u )τ sail + FCa (τ − τ sail )) − BH , τ ≥ τ sail Forward price ($/bbl) Loaded spot price ($/bbl) Cost of backhaul trip ($/bbl) Bunker price ($/mt) Fuel consumption (mt/day/bbl) Ship timecharter ($/bbl/day) 59 (4.2) The exercise profit consists in maximizing ω over all possible times-to-maturity: Ω(t , F (τ )) = max ω (τ ) (4.3) τ ≥τ sail Table 6. Ship and route parameters Unit Typical value Unit Location X Nm Ship1 Time t days Type Typical value Very Large Crude Carrier (VLCC) Cargo size Distance d Cargo DWT mt 300 000 mt 4535 Nm Speed u knots 15 knots Brent Nm 270 000 mt Sullom Voe – LOOP Route mt Fuel consumption sailing: FC(u) mt/day 87.5 mt/day (laden) 74 mt/day (ballast) 85 mt/day (pumping) 15 mt/day (anchor) anchor: FCa mt/day Barrel factor bbl/mt 7.578 bbl/mt Timecharter price H USD/day Loading port X0 Sullom Voe Delivery port XP VLCC average timecharter equivalent (Baltic Exchange) LOOP Loading price S0 USD/bbl Dated Brent 10-21 days Delivery price F(t, τ) USD/bbl LLS forward curve Loading delay τload Days 15 days IFO price B USD/mt Fuel Oil 3.5% CIF NWE (Platts) It should be noted that if the location of the ship X is the loading port, then Ω is the arbitrage profit from that port and if it is positive, the arbitrage is said to be open. Furthermore, if the loading and destination ports are the same and Ω is positive, then there is a floating storage opportunity at that port and Ω is the profit that can be made from it. This profit is also riskless – at least market-wise – considering that the profit is locked in by selling the cargo forward. Valuing the expected profit of the voyage When the tanker is loaded, the arbitrage profit Ω can be locked in without risk. However, the large number of optionalities available to the trader throughout the voyage means that the expected profit is sometimes higher. 1 Typical values correspond to the modern double-hull VLCC from Clarksons (2009) 60 Let us define V (t , F (τ )) as the expected profit from the cargo when the tanker is at the location X (t ) and the forward curve is given by F (τ ) . This is a real option value as detailed in Dixit and Pindyck (1994). When the ship is at a location X (t ) , and the maximum exposure time T has not been exceeded, the trader has two choices: • either “exercise” and sell the cargo forward, thereby earning the exercise profit Ω(t , F (τ )) • or choose to continue speculating during a time dt without exercising. The expected profit is: VC (t , F (τ )) = EF (τ ) [V (t + dt , F (τ ) + dF (τ )) ] − g (t )dt (4.4) This gives the continuation value VC . The forward curves are evolved during the time period dt using the two-factor model from Part 2. Hence the expected profit at location X (t ) given the forward curve F (τ ) is V (t , F (τ )) = max [ Ω(t , F (τ )), VC (t , F (τ ))] (4.5) We notice that we always have V ≥ Ω because the possibility of obtaining Ω is included in V. If the maximal exposure time t = T is reached, the cargo must be sold and V (t , F (τ )) = Ω(t , F (τ )) . This value function V contains all the information needed to evaluate and run the physical trade: • the value V (0, F0 (τ )) is the expected profit from following the optimal trading strategy • the a priori risk of the strategy and its exposure to the principal risk factors can be evaluated through V, as seen in Section 3.5. • at a date t, given the forward curve F (τ ) , compare the exercise value Ω(t , F (τ )) and the continuation value VC (t , F (τ )) . o If Ω ≥ VC then the delivery of the cargo should be specified. The optimal date at which to deliver it is obtained from the calculation of Ω o If VC > Ω then the ship should continue sailing or anchoring without specifying when delivery will take place. Simplification in the case of a two-factor model If the dynamics of the forward curve are described by a simple two-factor model as described in Part 2, the forward curve F (τ ) can be expressed in terms of the factor values f1 and f 2 and the initial forward curve F0 (τ ) : log Ft (τ ) = log F0 (t + τ ) +ψ (t, τ ) + u1 (τ ) f1 (t ) + u2 (τ ) f 2 (t ) 61 (4.6) Therefore the functions Ω and V only depend on the values of f1 , f 2 and time t since the beginning of the trade: Ω(t , F (τ )) = Ω(t , f1 , f 2 ) V (t , F (τ )) = V (t , f1 , f 2 ) (4.7) Optimal stopping formulation Determining V can alternatively be seen as an optimal stopping problem. The value function can equivalently be written as τ V (t , f1 , f 2 ) = max E − ∫ g ( s )ds + Ω(τ , f1 (τ ), f 2 (τ )) | f1 (t ) = f1 , f 2 (t ) = f 2 τ ∈ST[ t, T ] t (4.8) where ST[ t, T ] is the set of all stopping times in [t , T ] . The optimal stopping time corresponds to the time when V becomes equal Ω , i. e. τ * = min (4.9) and the initial expected profit from the trade is τ V = E − ∫ g ( s )ds + Ω(τ * , f1 , f 2 ) 0 * (4.10) 3. Solution methods Solving an optimal stopping problem such as the one that has been formulated for the floating storage trade is akin to calculating the value of an American option. A number of numerical methods have been suggested to this effect and we will review some of them here. Dynamic programming The conceptually simplest method of solving an American option problem is by dynamic programming. Discretizing time into dates t0 = 0, t1 . t N = T , the value at date ti can be written as 62 V (t N , f1 , f 2 ) = Ω(t N , f1 , f 2 ) VC (ti , f1 , f 2 ) = E [V (ti +1 , f1 + ∆f1 , f 2 + ∆f 2 ) ] − g (ti )∆t (4.11) V (ti , f1 , f 2 ) = max [ Ω(ti , f1 , f 2 ), VC (ti , f1 , f 2 )] The expectation in the calculation of VC (ti , f1 , f 2 ) is calculated using the transition probabilities of f1 and f 2 . When a binomial distribution is assumed this yields the binomial tree method for American options, as detailed in Clewlow and Strickland (1998). Using the two-factor model presented here we can evaluate it using transition probabilities. 1 2 The factor value space is discretized into a N1 × N 2 rectangular grid: ( f11 . f N1 ) × ( f12 . f N2 ) . The expectation is evaluated numerically using the previously calculated values of V (t j +1 , f k1 , f l 2 ) and the joint probability density of (df 1 , df 2 ) : 1 E f 1 , f 2 V (t j +1 , f k1 + df 1 , fl 2 + df 2 ) ≈ NF kl pk 'l ' = N1 N 2 ∑∑ p k '=1 l ' =1 V (t j +1 , f k1' , f l 2 ) ' k 'l ' ( f − f − ( µ1 − α1 f k1 )∆t ) 2 ( f l 2 − f l 2 − ( µ 2 − α 2 f l 2 )∆t ) 2 exp − −' 2 2πσ 1σ 2 ∆t 2σ 12 ∆t 2σ 2 ∆t 1 k' 1 1 k (4.12) NF = ∑ pk 'l ' k ',l ' Partial Differential Equation In continuous time, the dynamic programming formulation for V combined with Ito’s formula yields a partial differential equation for V . Let us assume that VC > Ω , i. e. we are in the continuation region, such that V = VC . We develop V using Ito’s formula: V (t + dt , f1 + df1 , f 2 + df 2 ) = V + ∂V ∂V ∂V 1 ∂ 2V dt + df1 + df 2 + ∑ dfi df j ∂t ∂f1 ∂f 2 2 1≤i , j ≤ 2 ∂fi ∂f j ∂V ∂V ∂V 1 2 ∂ 2V 1 2 ∂ 2V (4.13) =V + + ( µ1 − α1 f1 ) + (µ2 − α 2 f 2 ) + σ1 + σ 2 2 dt ∂f1 ∂f 2 2 ∂f12 2 ∂f 2 ∂t ∂V ∂V + σ 1dW1 + σ 2 dW2 ∂f1 ∂f 2 Using the definition of V and VC we find that V (t , f1 , f 2 ) = VC (t , f1 , f 2 ) = E [V (t + dt , f1 + df1 , f 2 + df 2 )] − g (t )dt 63 (4.14) ∂V ∂V ∂V 1 2 ∂ 2V 1 2 ∂ 2V + ( µ1 − α1 f1 ) + (µ2 − α 2 f 2 ) + σ1 + σ 2 2 = g (t ) ∂t ∂f1 ∂f 2 2 ∂f12 2 ∂f 2 (4.15) This is valid for (t , f1 , f 2 ) in the continuation region such that V (t , f1 , f 2 ) > Ω(t , f1 , f 2 ) . For (t , f1 , f 2 ) in the exercise region we have V (t , f1 , f 2 ) = Ω(t , f1 , f 2 ) . If we define S * (t ) as the continuation region at time t, this entails the following boundary conditions on V V (t , f1 , f 2 ) = Ω(t , f1 , f 2 ), ∀( f1 , f 2 ) ∈ ∂S * (t ) (continuity) ∇V (t , f1 , f 2 ) = ∇Ω(t , f1 , f 2 ), ∀( f1 , f 2 ) ∈ ∂S * (t ) (smooth pasting) (4.16) Using these equations for V we can calculate V using finite differences or a semi-analytical formulation. Semi-analytical solution Following Albanese and Campolieti (2006), the partial differential equation for V can be solved in closed form if we assume the boundary of the continuation region to be known. In Appendix 7 we show that the solution of (4.15) can be written as T V (t , f ) = ∫ p( f ', T ; f , t )(Ω(T , f ') − G (t , T ))df ' + ∫ ∫ ℝ2 p ( f ', s; f , t )(−ψ ( s, f '))df ' ds t ℝ2 \ S * ( s ) (4.17) = Veur (t , f ) + Vearly (t , f ) where T is the maximum exposure time, and 0 f ∈ S * (τ ) ψ (t , f ) = ∂Ω + LΩ f ∉ S * (τ ) ∂τ ∂Ω ∂Ω 1 2 ∂ 2 Ω 1 2 ∂ 2 Ω L Ω = ( µ1 − α1 f1 ) + ( µ2 − α 2 f 2 ) + σ1 + σ2 − g (t ) ∂f1 ∂f 2 2 ∂f12 2 ∂f 22 (4.18) The continuation region S * (t ) is defined as S * (t ) = (4.19) The boundary ∂S * (t ) of this domain has to be determined for each date t. We write it as a function of the second factor ∂S * (t ) = ( f1* (t , f 2 ), f 2 ), f 2 ∈ ℝ 64 (4.20) such that the equation to be solved by f1* (t , f 2 ) is V (t , f1* (τ , f 2 ), f 2 ) = ∫ p( f , f , T ; f ' 1 ' 2 * 1 (t , f 2 ), f 2 , t )(Ω(T , f ') − G (t , T ))df ' ℝ2 T +∞ +∫ +∞ ∫∫ p1 ( f1' , f 2' , s; f1* (t , f 2 ), f 2 , t )(−ψ ( s, f '))df ' ds (4.21) t −∞ f1* ( s , f 2 ) = Ω(t , f1* (t , f 2 ), f 2 ) In Appendix 7 we detail the numerical procedure used to find this exercise boundary, which can be found recursively beginning at t = T . Monte Carlo simulation methods The methods discussed above, while suitable for a two-factor model, become impractical if the number of factors is higher, for example if several ports are being considered. In this case a Monte Carlo method should be employed. Monte Carlo methods are not ideally suited to American option problems, because of their backward-recursion properties. However, Longstaff and Schwartz (2001) suggest a least-squares Monte Carlo method with projection of the value function onto a small basis, allowing for efficient pricing of American options. A similar method could be employed in this case. 4. Analytical properties of the solution We will examine some of the properties of the expected profit function V using the analytical expression obtained above. We will decompose the solution as follows: V (0, f ) = ∫ ℝ T p ( f ', T ; f , 0)(Ω(T , f ') − G (0, T ))df ' + ∫ 2 ∫ 2 p ( f ', s; f , 0)(−ψ ( s, f '))df ' ds * 0 ℝ \S (s) (4.22) = Ω(0, f ) + EPdrift + EPconvexity + EPearly where the excess profit components EPdrift , EPconvexity and EPearly are defined as EPdrift = Ω(T , Ef1 (T ), Ef 2 (T )) − G (T ) − Ω(0, f1 , f 2 ) EPconvexity = Veur (T , f1 , f 2 ) − (Ω(T , Ef1 (T ), Ef 2 (T )) − G (T )) EPearly = Vearly (T , f1 , f 2 ) 65 (4.23) Model parameter dependence We want to examine the dependence of the value function on the model parameters α k , σ k and µ k . Let us consider first the drift component EPdrift . Its value does not depend on the volatilities of the factors. Remembering that Ef k (T ) = f k e−α k T + µk (1 − e−α T ) αk (4.24) k we can derive its dependence on the drift and on the model parameters: ∂EPdrift 1 − e−α k T ∂Ω = (T , Ef1 (T ), Ef 2 (T )) ∂µ k α k ∂f k ∂EPdrift ∂Ef k (T ) ∂Ω = (T , Ef1 (T ), Ef 2 (T )) ∂α k ∂α k ∂f k (4.25) ∂EPdrift =0 ∂σ k The convexity component can be written as ∫ p( f ', T ; f , 0)(Ω(T , f ') − Ω(T , Ef (T ), Ef EPconvexity = 1 ℝ ∫ p( f = ' 1 (T )))df ' ( ) − Ef1 (T ), f 2' − Ef 2 (T )) Ω(T , f1' , f 2' ) − Ω(T , Ef1 (T ), Ef 2 (T )) df ' ℝ2 ∫ p( f , f ) ( Ω(T , Ef (T ) + f , Ef = 2 2 ' 1 ' 2 1 ' 1 2 (4.26) ) (T ) + f 2' ) − Ω(T , Ef1 (T ), Ef 2 (T )) df ' ℝ2 (in the last line we change variables from f ' to f ' + Ef (T ) ). We want to show that this convexity premium does not depend on the expected value of the factors. ∂EPconvexity ∂Ef k = This term is zero if ∫ p( f , f ' 1 ℝ2 ' 2 ∂Ω ∂Ω ) (T , Ef1 (T ) + f1' , Ef 2 (T ) + f 2' ) − (T , Ef1 (T ), Ef 2 (T )) df ' (4.27) ∂f k ∂f k is at most quadratic in the factors. In the general case we can write ∂Ω ∂Ω ∂Ω '2 ( Ef + f ' ) − ( Ef ) = f '⋅∇ + O( f ) ∂f k ∂f k ∂f k such that: 66 (4.28) ∂EPconvexity ∂Ef k −α 2T 1 − e −α1T 2 1− e ≤ C ∫ p ( f1' , f 2' )( f1'2 + f 2'2 )df ' = C σ 12 +σ2 α1 α2 ℝ2 (4.29) For sufficiently small values of T this term is small, of the order O (σ 12T ) . Hence we have established that the excess profits coming from convexity are indeed independent of the expected values of the factors, and therefore also of the drifts. Early exercise in a backwardated market As we will see in Section 4.6, it is generally optimal to sell the cargo immediately when the forward curve net of freight is in backwardation. We will show this result here. We assume that the loading and delivery port are the same, such that the trade is purely a floating storage trade. The forward curve net of freight is in net backwardation if F0 (τ ) − ( H + FCa )τ < F0 (0) (4.30) In this case it is optimal to sell the cargo spot, such that t Ω(t ) − ∫ g ( s )ds − Ω(0) = F0 (t ) − ( H + FCa )t − F0 (0) < 0 (4.31) 0 The initial expected excess profit is, as seen in equation (4.10), τ V (0, 0, 0) = E − ∫ g ( s )ds + Ω(τ * , f1 , f 2 ) 0 * (4.32) such that ( ) V − Ω = E F0 (τ * ) − ( H + FCa )τ * − F0 (0) + E F0 (τ * ) exp(u1 (0) f1 (τ * ) + u2 (0) f 2 (τ * )) − 1 (4.33) Conditional on τ * = τ (independent of the values of f1 and f 2 , we then find that given the distributions of f1 (τ ) and f 2 (τ ) , ( ) E F0 (τ * ) exp(u1 (0) f1 (τ * ) + u2 (0) f 2 (τ * )) − 1 = * * 2 u (0) 2 σ 12 (1 − e −α1τ ) u2 (0) 2 σ 2 (1 − e −α 2τ ) F0 (τ * ) exp u1 (0) E[ f1 (τ * )] + u2 (0) E[ f 2 (τ * )] + 1 + − 1 2α1 2α 2 We now introduce the initial slope of the curve a = ∂F0 ∂τ 67 (4.34) and consider only small values of τ * , then: τ =0 a−g 1 2 V − Ω = F0 (0) + u1 (0) µ1 + u2 (0) µ 2 + (u1 (0)2 σ 12 + u2 (0)σ 2 ) τ * + O(τ *2 ) 2 F0 (0) (4.35) a−g 1 2 + u1 (0) µ1 + u2 (0) µ 2 + (u1 (0) 2 σ 12 + u2 (0)σ 2 ) < 0 F0 (0) 2 (4.36) Hence, if there is no value to exercising later, such that V = Ω . The other parameters being fixed, this can always be achieved for a strong enough net backwardation. 5. Profit and risk The calculation of the functions V and defines a physical trading strategy that can be applied in practice. In order to assess how interesting this strategy is, we would like to assess a priori its expected return and risk. Furthermore we would like to assess the dependence of the profits on the different risk factors, to define a financial hedging strategy using futures or options. Expected and realized profit on the trading strategy As discussed above, the values V (t = 0, f1 = 0, f 2 = 0) and Ω(t = 0, f1 = 0, f 2 = 0) are respectively the expected profit and arbitrage profit that can be obtained on the initial date. These numbers are expressed in US dollars per barrel ($/bbl) for crude oil or US dollars per gallon ($/gal) for gasoil. When the trading strategy is executed the realized profit is not necessarily equal to the expected profit, given that the distribution of forward curves is stochastic. The realized profit of the trip is, in the simple case of a single port, t* W = ∫ − g (t )dt + Ω(t * , f1 (t * ), f 2 (t * )) (4.37) 0 where t * is the exercise date. This can be calculated a posteriori to get the realized profit. But seen at t = 0 this is a random variable with a certain distribution. Its expected value is V: V (t = 0, f10 , f 20 ) = E W (τ * , f1 (τ * ), f 2 (τ * )) | f1 (t = 0) = f10 , f 2 (t = 0) = f 20 (4.38) where τ * is the optimal stopping time, which is a random variable depending on the realized values of f1 and f 2 . Expected risk and Sharpe ratio There is no market risk tied to the arbitrage profit Ω, because the cargo is sold forward and the profit is fixed at the moment the decision is taken. However there is financial risk tied to the physical trading strategy with 68 expected profit V: the forward curves will change before the decision to deliver the cargo is taken. This risk is reflected in the distribution of the realized profit W. We have seen that this distribution is centered on V: V (t = 0, f10 , f 20 ) = E W (τ * , f1 (τ * ), f 2 (τ * )) | f1 (t = 0) = f10 , f 2 (t = 0) = f 20 (4.39) Furthermore, at exercise, t* t* W (t , f1 , f 2 ) = ∫ − g (t )dt + Ω(t , f1 , f 2 ) = ∫ − g (t )dt + V (t * , f1 , f 2 ) * * 0 (4.40) 0 We define the process U representing the expected profit and loss (P&L) on the trade at time t by t U (t , f1 , f 2 ) = ∫ − g ( s )ds + V (t , f1 , f 2 ) (4.41) 0 The value of this process at exercise equals W, the realized P&L of the trade. To find its distribution we differentiate U using Ito’s formula: ∂V ∂V ∂V 1 2 ∂ 2V 1 2 ∂ 2V ∂V ∂V dU = − g (t ) + + ( µ1 − α1 f1 ) + (µ2 − α 2 f 2 ) + σ1 + σ 2 2 dt + σ 1dW1 + σ 2 dW2 2 ∂t ∂f1 ∂f 2 2 ∂f1 ∂f 2 ∂f1 ∂f 2 2 (4.42) ∂U ∂U σ 1dW1 + σ 2 dW2 = ∂f1 ∂f 2 We find that the process U has zero drift. This reflects the fact that V was correctly priced initially. The instantaneous volatility of U over a time period dt is 1/2 ∂U 2 ∂U 2 σ U = σ 1 + σ 2 ∂f1 ∂f 2 ( = σ 1 ∆1 )( 2 ) 1/2 2 + σ 2 ∆2 (4.43) where ∆k = ∂V ∂f k (4.44) is the delta of the value function with respect to factor k. Furthermore, the distribution of U given the stopping time τ * is τ* τ* ∂U ∂U U (τ , f1 , f 2 ) = U (0, f , f ) + ∫ (t , f1 (t ), f 2 (t ))σ 1dW1 (t ) + ∫ (t , f1 (t ), f 2 (t ))σ 2 dW2 (t ) (4.45) ∂f1 ∂f 2 0 0 * 0 1 0 2 We can calculate the first moments of U: 69 E[U (τ * , f1 , f 2 )] = U (0, f10 , f 20 ) = V (0, f10 , f 20 ) (4.46) 2 2 τ * ∂V τ * ∂V 2 U (τ , f1 , f 2 ) = σ E ∫ Var (t , f1 (t ), f 2 (t )) dt + σ 2 E ∫ (t , f1 (t ), f 2 (t )) dt (4.47) 0 ∂f1 0 ∂f 2 * 2 1 The variance depends on the stopping time and can best be evaluated through Monte Carlo simulation, simulating the paths of ( f1 , f 2 ) and using the value function already calculated. If we make the simplifying assumption that the deltas of the value function are constant, the variance can be approximated as: ( ) 2 2 Var[W ] = Var U (τ * , f1 , f 2 ) ≈ σ 12 ∆1 + σ 2 ∆ 2 E[τ * ] 2 (4.48) Alternatively, the complete distribution of W can be evaluated using Monte Carlo simulation. It should be noted that this Monte Carlo simulation is simpler than the least squares Monte Carlo technique used for finding the optimal stopping time. If we know the expected profit and the standard deviation of the realized profit, we can calculate the annualized Sharpe ratio of the strategy a priori: Expected profit Std. deviation 1 V = 1/ 2 * 2 2 E[τ * ]1/ 2 τ ∂V τ * ∂V 2 σ 12 E ∫ (t , f1 (t ), f 2 (t )) dt + σ 2 E ∫ (t , f1 (t ), f 2 (t )) dt 0 ∂f1 0 ∂f 2 V ≈ 1/ 2 2 2 σ 12 ∆1 + σ 2 ∆ 2 E[τ * ] 2 SR = ( (4.49) ) 6. Results Arbitrage results In this section we present the results from the calculation of the function Ω at different dates. This function, evaluated at trade initiation time (t = 0), gives the arbitrage profit that can be obtained from the shape of the forward curve at the current date. By studying its dependence on the factor values f1 and f 2 , we can also evaluate its dependence on the level and slope of the curve. 70 20 A rb profit (Ω ) Spread Freight cost 15 $/bbl 10 5 0 -5 06/07/07 09/15/07 12/24/07 04/02/08 07/11/08 10/19/08 Date 01/27/09 05/07/09 08/15/09 11/23/09 Figure 17. Arbitrage profit per barrel on the Sullom Voe-LOOP route 250 Time to delivery (days) 200 150 100 50 0 06/07/07 09/15/07 12/24/07 04/02/08 07/11/08 10/19/08 Date 01/27/09 05/07/09 08/15/09 11/23/09 Figure 18. Time to delivery for static arbitrage on the Sullom Voe – LOOP route 71 160 BFOE Sullom Voe spot LLS LOOP spot 140 Price ($/bbl) 120 100 80 60 40 20 06/07/07 09/15/07 12/24/07 04/02/08 07/11/08 10/19/08 01/27/09 05/07/09 08/15/09 11/23/09 Figure 19. Spot prices of BFOE Sullom Voe (blue) and LLS LOOP (green) 10 LLS LOOP forward curve slope Slope ($/bbl/month) 5 0 -5 -10 -15 06/07/07 09/15/07 12/24/07 04/02/08 07/11/08 10/19/08 01/27/09 05/07/09 08/15/09 11/23/09 Figure 20. Slope of the LLS forward curve, in US Dollars per barrel per month, measured on the front two month contracts 72 On each day in the sample period (August 2007 to October 2009) we calculate the arbitrage profit Ω that can be obtained from the observed forward curve and freight prices on that date. The arbitrage profit is calculated as Ω = max ω (τ ) where τ ≥τ sail τ sail = d ( X 0 , X P ) / u ω (τ ) = F (τ ) − S0 − Hτ − B ⋅ ( FC (u )τ sail + FCa (τ − τ sail )) − BH , τ ≥ τ sail (4.50) The arbitrage profit is obtained by buying BFOE crude at Sullom Voe on the trade initiation date and delivering it into LOOP in the optimal time τ * , sailing at speed u to get there, and anchoring up for a time τ − τ sail to wait for delivery. The cargo is bought at the spot price S0 and sold forward at the price F (τ * ) . We decompose this arbitrage profit into a geographical spread Sp(τ ) = F (τ ) − S0 and a freight cost Fc , Fc(τ ) = Hτ + B ⋅ ( FC (u )τ sail + FCa (τ − τ sail )) + BH (4.51) The physical arbitrage is said to be open when Ω > 0 , i. e. Sp(τ * ) > Fc(τ * ) : the spread that can be earned on the crude oil is higher than the cost of transportation and storage. In the opposite case it is said to be closed. When the arbitrage window is open it is profitable to ship a cargo of oil on the considered route. We present the results for the Sullom Voe-LOOP route in Figure 17. The geographical spread is shown in green, the freight cost in red, and the net arbitrage profit Ω in blue. We observe that the arbitrage window is open during large parts of the time period under consideration. In Figure 18 we show the value of τ * , the time between the current date and the optimal date to exercise the option of delivering the cargo. We observe that the large profits from late 2008 and early 2009 came from the large opportunities in floating storage created by a steep contango and low timecharter rates. In Figure 19 we show the spot prices of crude oil at the loading and delivery ports. In Figure 20 we show the slope of the forward curve at the delivery port. Figure 19 confirms what makes this physical arbitrage possible: the spread in spot prices between European and American crude. However, the arbitrage profits seem to be uncorrelated with the general level of crude prices. This stems from the fact that international crude prices largely move together, partly because of such arbitrage activity. There does, however, seem to be some relation between the forward curve slope and the arbitrage profit. This is witnessed in Figure 21 where we regress the arbitrage profit on the forward curve slope. The relationship is stronger for a steeper contango. In Figure 22 we present data collected from different research reports on the actual crude oil in floating storage worldwide alongside the optimal time to delivery for the arbitrage trade. We see that there is a substantial increase in the amount of crude oil stored at sea starting in October 2008. This coincides with the appearance of floating storage opportunities according to our model. Furthermore, the short disappearance of floating storage opportunities according to our model in June 2009 was accompanied by a clear downward trend in the number of tankers storing crude in the Goldman Sachs and Gibson Research data. 73 The same analysis can be performed for different markets and different routes. As a point of comparison we present the results for the gasoil arbitrage between Europe and the United States. The product being traded is No. 2 fuel oil, also known as gasoil or heating oil. The loading port is the Amsterdam-Rotterdam-Antwerp (ARA) region, Europe’s major refining hub. The destination is New York harbor (NYH), which is the main delivery point of refined products on the east coast of the United States. Details of the route and cargo are presented in Appendix 8. The arbitrage profit, decomposed as described above, is presented in Figure 23. The optimal delivery time is presented in Figure 24 and these results should be compared to the spot prices in Figure 25 and the forward curve slope in Figure 26. We note that the arbitrage window is open less frequently than was the case for crude oil and the spread has been negative on occasions, making the inverse arbitrage (U. S. to Europe) interesting. However, there have been significant floating storage opportunities since the end of 2008 as witnessed in Figure 24, and these profits have been very interesting: around 20 cents per gallon for a gallon costing less than 2 dollars. 16 Data Linear regression 14 12 Arbitrage profit (Ω ) 10 y = 3.35 + 0.91x Rsq = 0.30 8 6 4 2 0 -2 -4 -2 0 2 4 Forward slope ($/bbl/month) 6 8 10 Figure 21. Relationship between the forward curve slope at the delivery port (LLS LOOP) and the arbitrage profit 74 180 350 Morgan Downey (crude + resid) IEA/Goldman Sachs 300 Gibson research 140 250 Time to delivery 120 100 200 80 150 60 100 40 Time to delivery (days) Crude floating storage (mn bbl) 160 50 20 0 0 6/1/2008 9/9/2008 12/18/2008 3/28/2009 7/6/2009 10/14/2009 1/22/2010 5/2/2010 Figure 22. Crude oil in floating storage worldwide (left axis) and optimal time to delivery of the floating storage trade (right axis). Sources: IEA/Goldman Sachs Global ECS Research, Gibson Research, Morgan Downey A rb profit (Ω ) Spread Freight cost 0.5 0.4 0.3 0.2 $/gal 0.1 0 -0.1 -0.2 -0.3 -0.4 -0.5 02/23/04 07/07/05 11/19/06 04/02/08 08/15/09 12/28/10 Date Figure 23. Arbitrage profit per gallon for gasoil trade between ARA and NYH 75 600 Time to delivery (days) 500 400 300 200 100 0 02/23/04 07/07/05 11/19/06 04/02/08 08/15/09 12/28/10 Date Figure 24. Time to delivery for arbitrage on the ARA-NYH route Gasoil ARA spot Heating oil NYH spot 4.5 4 Price ($/gal) 3.5 3 2.5 2 1.5 1 02/23/04 07/07/05 11/19/06 04/02/08 08/15/09 12/28/10 Figure 25. Spot prices of No. 2 fuel oil at Amsterdam-Rotterdam-Antwerp (blue) and New York harbor (green), in US Dollars per gallon 76 Heating oil NYH forward curve slope 0.08 0.06 0.04 Slope ($/gal/month) 0.02 0 -0.02 -0.04 -0.06 -0.08 -0.1 -0.12 02/23/04 07/07/05 11/19/06 04/02/08 08/15/09 12/28/10 Figure 26. Slope of the heating oil forward curve at New York Harbor, in US Dollars per gallon per month Expected and excess profits The results presented for were static arbitrage results. We now consider the optimal trading strategy presented in Section 4.2. This trading strategy yields a value function V which is the expected profit of the physical trading strategy. These results are obtained using the semi-analytical formulation presented in Section 4.3, calculating the exercise boundary numerically as described in Appendix 7. The two-factor model used is calibrated on the crude oil futures market as described in Section 2.4, and we make the assumption that drifts are zero: the expected spot price is therefore equal to the forward price. Furthermore, trades are limited to a maximal exposure time T equal to 100 days. We study the shape of and V with the initial date set to December 18, 2008. As seen in Figure 27 the and V as functions of the factor values f1 and f 2 at forward curve on that date was in contango. We plot trade initiation. 77 60 F (τ) 0 100 F (τ) - Fc(τ) (net of freight) 0 80 55 60 Profit ($/bbl) F(τ) ($/bbl) 50 V 40 20 45 0 -20 40 -40 0.06 35 0 50 100 150 τ (days) 200 250 0.04 0.02 0 -0.02 -0.04 -0.06 0 -0.5 -1 1 0.5 300 f f 2 1 Figure 28. Ω and V as functions of the factors f1 and f2 at trade initiation (December 18, 2008) Figure 27. Forward curve and forward curve net of freight on December 18, 2008 V(f1,0), tau = 100 days V(0,f2), tau = 100 days 80 1 12 1 0.9 60 0.9 10 V Veur 0.8 0.8 Vearly 40 0.7 8 0.7 Profit ($/bbl) Profit ($/bbl) Ω 0.6 20 0.5 Ω 0 0.4 -40 0.1 -0.8 0 24 0.3 2 V early 0.2 0.2 f* 1 -0.6 29 -0.4 34 -0.2 41 0 f1 0.2 49 59 F1 level ($/bbl) 0.4 70 0.6 84 0.5 4 0.4 V V eur 0.3 -20 0.6 6 0.8 0 0.1 -0.06 -0.04 -0.02 0.02 0.04 0.06 101 0 +6.04 +5.67 +5.30 +4.93 +4.56 Fwd curve slope ($/bbl/month) +4.19 +3.82 0 f2 Figure 29. Cross-section of Ω and V at trade initiation as a function of f1 (left) and f2 (right) 78 Excess profit from continuation, V - Ω, τ = 100 days Profit ($/bb l) +4.01 1 0.05 +4.19 0.9 0.04 4 +4.38 0.03 0.8 3.5 +4.56 0.02 0.7 3 +4.74 0.01 0.6 2.5 +4.93 0 * f2 Fwd curve slope ($/bbl/month) ∂S (T) 0.5 -0.01 2 +5.11 +5.30 0.4 -0.02 1.5 +5.48 0.3 -0.03 1 +5.67 -0.04 0.2 0.5 +5.85 0 0.1 -0.05 0.1 0.2 0 -0.6 0.3 -0.4 0.4 -0.2 0.5 29 34 41 0.6 0 f1 0.7 0.2 0.8 0.4 0.9 0.6 49 59 F1 level ($/bbl) 70 1 0 84 Figure 30. Expected excess profit from continuation V - Ω at trade initiation, in USD/bbl, and exercise boundary (red line) In Figure 28 and Figure 29 we show the dependence of V and on f1 and f 2 . The values corresponding to the initial forward curve are V (0, 0, 0) and Ω(0, 0, 0) , valued respectively at 6.55 $/bbl and 5.81 $/bbl. The other values correspond to a forward curve that has been shocked by the factor values f1 and f 2 . We can see that, predictably, a positive parallel shift ( f1 > 0) yields a higher expected profit. The slope with respect to the second factor is lower. We can also see that V and are the same at the maximal exposure time: this is the terminal condition that we impose. Furthermore, at trade initiation V is higher than Ω , and more so for low values of Ω . Thus there is value to keeping the options open. For negative values of Ω it is still possible to have positive values of V: there is a chance that prices will rebound enough to yield a profit during the trade period. Figure 29 decomposes the value function V into two components: the European exercise value Veur and the early exercise premium Vearly . The European exercise value corresponds to the expected profit that would be earned if the cargo was held until the maximal exposure time T (100 days in this case), and then sold into the market. This value largely depends on the drifts of the factors. The mean-reverting model has a large impact in this respect. When the value of f1 is negative, it is expected to increase, which pushes the expected value up compared to the arbitrage value. When the value of f1 is positive, its expected value is lower, pushing the expected value down. 79 This is counteracted by the early-exercise premium, which is positive for high values of the first factor and for high absolute values of the second factor. This corresponds to situations where it is close to optimal to exercise. The decision to exercise is made based on the difference between the expected profit V and the exercise profit Ω . We plot this difference in Figure 30. The darkest zone, where V = Ω , is the exercise region. If ( f1 , f 2 ) falls in this zone it is optimal to specify delivery of the cargo and harvest the profit Ω . Outside this region it is optimal to continue sailing and delay the decision about delivery time until later. The excess profit is seen to depend on the shape of the forward curve through the factor values f1 and f 2 . The excess profit is seen to be highest when the first factor is lowest: because it is mean-reverting, keeping the position open gives more upside exposure than downside exposure. Dependence of expected profits on model parameters The results above are presented for a two-factor model that has been calibrated on the crude oil market as detailed in Part 2. We have seen that the interpretation of the excess profits is linked to the model parameters α , σ and µ which determine the distribution of possible forward curves. It is therefore interesting to examine the dependence of the profits on the values of these parameters. We vary the parameters within reasonable ranges around the reference values that have been used before, and plot the dependence of V, Veur , Vearly and Ω on these parameters. The trade initiation date is December 18, 2008. The results are presented in Figure 31. The dependence on the mean-reversion parameter is rather weak compared with the dependence on the other parameters. This can be explained by the fact that the maximal exposure time, 100 days, is rather short, and that the excess profit we are considering is taken at f10 = f 20 = 0 , such that the expected value of the factors is not affected by the mean-reversion parameter. The dependence on the volatilities of the factors is very strong, with a doubling of σ 1 from 26% to 52% taking the excess expected profit from 0.68 $/bbl to 2.19 $/bbl. The effect is larger in the first factor because the magnitude of the excess profits coming from the first factor are much larger. But in relative terms, doubling σ 2 from 1.86% to 3.72% takes the excess profit attributable to the second factor, i. e. the difference between the profit for a σ 2 larger than zero and the profit for σ 2 = 0 , from 7.78 c/bbl to 26.14 c/bbl, which is a significant increase. The effect of the drift parameters µ1 and µ 2 is to change the expectations about what the forward curve will look like in the future. In particular, a negative value for µ1 means that the trader is taking a sharply negative view on the future level of prices. In that case it is more interesting to exercise early to take the profits given the current level of prices. A positive value for µ1 is a positive view on levels and it will be preferable to wait to take advantage of rising prices. A negative value for µ 2 corresponds to a view of a sharper contango, which is beneficial to the trade, while more backwardation ( µ 2 > 0 ) is detrimental. 80 Figure 31. Dependence of expected profits on the model parameters Dependence of profits on α Dependence of profits on α 2 1 7 7 6 6 Ω V Veur 5 Profit ($/bbl) 4 Profit ($/bbl) 5 Vearly 3 Reference 2 V Veur Vearly 3 Reference 2 1 1 0 -1 Ω 4 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 0 1.8 α1 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 α2 Dependence of profits on α1. Reference is α1 = 0.84 yr-1 Dependence of profits on α2. Reference is α2 = 0.84 yr-1 Dependence of profits on σ 2 Dependence of profits on σ 1 7 9 8 6 7 5 Profit ($/bbl) 5 Profit ($/bbl) 6 Ω 4 V V eur 3 V early 2 Reference Ω 4 V Veur 3 Vearly Reference 2 1 1 0 -1 0 0.1 0.2 0.3 0.4 0.5 0.6 0 0 0.7 σ1 0.005 0.01 0.015 0.02 0.025 Dependence of profits on σ1. Reference is σ1 = 26% (annualized) Dependence of profits on µ Dependence of profits on µ 2 1 7 6 10 Ω 5 V V Profit ($/bbl) 8 Profit ($/bbl) 0.035 6 Ω V Veur 4 V 2 -0.2 -0.1 0 0.1 0.2 early 0.3 4 V eur early Reference 3 2 1 Reference -0.3 0.04 Dependence of profits on σ2. Reference is σ2 = 1.86% (annualized) 12 0 -0.4 0.03 σ2 0 -0.025 -0.02 -0.015 -0.01 -0.005 0.4 0 0.005 0.01 0.015 0.02 0.025 µ2 µ1 Dependence of profits on µ2. Reference is µ2 = 0 yr-1 Dependence of profits on µ1. Reference is µ1 = 0 yr-1 81 Dependence of profits on the ship speed In some circumstances, it can be beneficial for the trade to sail the ship slowly across the Atlantic in order to save on fuel costs. Intuitively, this will be especially useful when the forward curve is in a slight contango. The speed will affect profits in three ways: • A faster ship will be able to deliver its cargo earlier, which is important in a strong backwardation • A faster ship will be chartered for less time, such that its total time charter cost will be lower • A faster ship will consume more fuel. The fuel consumption function FC (u ) is approximately cubic in the speed u. In Figure 32 we examine the variation of the profits with the speed of the ship u for a trade beginning on August 13, 2008 and April 28, 2009. We notice that the speed has a small influence on profits, of the order for 10 c/bbl for a speed varying from 8 to 17 knots. The speed is fixed during the voyage. When the forward curve is in backwardation, there is incentive to deliver the cargo as soon as possible. A higher speed allows the trader to deliver the cargo earlier, but at the cost of higher fuel consumption. There is an optimal speed of around 13 knots yielding the best tradeoff. When the forward curve is in contango, the trade will involve some amount of floating storage at destination, such that fuel savings can be interesting. The excess profit V − Ω , however, is not affected by the speed. 11 1.6 1.5 Ω 1.4 10 Freight cost Spread Ω V Freight cost (Ω ) 9 1.3 Profit ($/bbl) Profit ($/bbl) Spread (Ω ) 1.2 1.1 1 0.9 8 7 6 0.8 5 0.7 8 9 10 11 12 u (knots) 13 14 15 4 8 16 9 10 11 12 13 u (knots) 14 15 16 17 April 28, 2009 (contango) August 13, 2007 (backwardation) Figure 32. Variation of profits with vessel speed (fixed during the voyage) on two different dates, when the forward curve was in backwardation (left) and contango (right) 82 Time series of expected profit, risk and Sharpe ratio For each week in the sample period, we perform the above calculations and derive: , the expected profit V and the excess profit V − Ω • The arbitrage profit • The expected risk, i. e. the standard deviation of W • The a priori Sharpe ratio We plot these values as a function of time. 15 Ω V V eur Vearly 10 Profit ($/bbl) 5 0 -5 -10 06/07/07 09/15/07 12/24/07 04/02/08 07/11/08 10/19/08 01/27/09 05/07/09 08/15/09 11/23/09 Figure 33. Arbitrage profit Ω and expected profit V, decomposed into Veur and Vearly, for weekly loading dates from August 2007 to August 2009 83 1.4 12 V-Ω 1.2 10 Standard deviation ($/bbl) Excess profit ($/bbl) 1 0.8 0.6 8 6 4 0.4 2 0.2 0 06/07/07 09/15/07 12/24/07 04/02/08 07/11/08 10/19/08 01/27/09 05/07/09 08/15/09 0 06/07/07 09/15/07 12/24/07 04/02/08 07/11/08 10/19/08 01/27/09 05/07/09 08/15/09 11/23/09 11/23/09 Figure 34. Expected excess profit V – Ω for different loading dates Figure 35. Expected standard deviation of realized profits. The zero values correspond to dates when exercise is immediate 0.7 40 35 SR on V - Ω 0.6 30 25 Sharpe ratio Exposure time (days) 0.5 20 0.4 0.3 15 0.2 10 0.1 5 0 06/07/07 09/15/07 12/24/07 04/02/08 07/11/08 10/19/08 01/27/09 05/07/09 08/15/09 11/23/09 Figure 36. Expected exposure time 0 06/07/07 09/15/07 12/24/07 04/02/08 07/11/08 10/19/08 01/27/09 05/07/09 08/15/09 11/23/09 Figure 37. A priori annualized Sharpe ratio1 of the excess profit V – Ω E[t * ] of the trade. The maximal exposure time is T = 100 days 1 The average long-term Sharpe ratio of the S&P500 is about 0.4. The Sharpe ratios in Figure 37 are calculated on profits over the riskless arbitrage profit . They are on the order of 6 when calculated over the risk-free rate. 84 1.4 V-Ω Excess profit V - Ω ($/bbl) 1.2 1 0.8 0.6 0.4 0.2 0 -2 0 2 4 6 Forward curve slope ($/bbl/month) 8 10 Figure 38. Excess expected profit V – Ω vs. Forward curve slope at delivery port In Figure 33 we plot the arbitrage profit and the expected profit V as a function of the trade initiation date. 나는. each week we examine the forward curve, spot price and shipping cost and determine what the arbitrage profit would be for a cargo loaded within the loading window τ load (15 days), as well as the expected profit V from loading the cargo and executing the optimal trading strategy. We see that these profits are always at least as great as the arbitrage profits. We have already studied the behavior of the time series of , so we will concentrate here on the excess profit, V – . We plot this expected excess profit as a function of time in Figure 34. Its value varies between 0 and 1 $/bbl, averaging 74 c/bbl in the period when the excess profit is positive. We note that the period under consideration can be separated in two: from 2007 to mid-2008 the continuation value is zero, while after the market crash in 2008 the excess profit jumps to values around 75 c/bbl. The first period corresponds to a backwardated forward curve, while the second period corresponds to a period of strong contango and low freight rates following the crisis. In Figure 38 we show the relationship between the forward curve slope at the delivery port and the excess profit from continuation. Consistently with what was proved in Section 4.4, we find that a forward curve in backwardation or in slight contango yields a zero excess profit, while all the positive excess profits are associated with a forward curve in contango. The standard deviation of the profits over the trade, presented in Figure 35, is significant, averaging 6.66 $/bbl in the period when the excess profit is positive. This is the risk associated with keeping the exposure to the forward curve open, and is accordingly zero when the cargo should be sold forward immediately, i. e. V = . The expected time over which this exposure is held E[t * ] is presented in Figure 36, and averages 27 days. 85 We note that the exposure time never reaches the maximal exposure time T that is set to 100 days here. Combining expected profit and risk we can calculate the annualized Sharpe Ratio associated with the strategy, which is presented in Figure 37. We consider this Sharpe ratio in excess of the riskless profit . It averages 0.41 during the period. In Section 4.6 we examine the detail of these time series and attempt to explain the appearance of excess profits. Realized profit and standard deviation The functions Ω(t , f1 , f 2 ) and V (t , f1 , f 2 ) define a physical trading strategy that can readily be put into practice. Given historical time series of the actual moves in the forward curve we can calculate the profits that would have been realized by following this strategy, and compare them to the expected profits and risks presented above. Statistics1 35 Realized profit W ($/bbl) Expected profit V ($/bbl) 95% C. I. for W E[W – V] 0.74 $/bbl Std[W – V] 6.21 $/bbl 20 SR on W 5.95 15 SR on W – 1.99 30 Profit ($/bbl) 25 10 5 0 -5 -10 -15 06/07/07 09/15/07 12/24/07 04/02/08 07/11/08 10/19/08 01/27/09 05/07/09 08/15/09 11/23/09 Figure 39. Expected profit V and realized profit W on different dates 1 Statistics are not calculated on the entire period, but on the period when V > 86 Statistics 120 Realized exposure time t* (days) * E[trealized − E[t * ]] * -1.5 days * Std[trealized − E[t * ]] Expected exposure time E[t*] (days) 15 days 95% C. I. for t 100 Exposure time (days) 80 60 40 20 0 06/07/07 09/15/07 12/24/07 04/02/08 07/11/08 10/19/08 01/27/09 05/07/09 08/15/09 11/23/09 Figure 40. Expected and realized exposure times t* On each week in the sample period, having calculated the functions Vt0 (t , f1 , f 2 ) and Ωt0 (t , f1 , f 2 ) with trade initiation date t0 , we execute the trading strategy defined by: • if Vt0 (0, 0, 0) > 0 , the trade is expected to be profitable, so initiate the trade by buying the cargo and chartering the vessel • At each time from trade initiation t > 0 , observe the forward curve F (t, τ ) and calculate the factor values using the orthogonality condition in (2.11), f k (t ) = τ max F (t, τ ) dτ F0 (τ ) ∫ u (τ ) log k 0 • (4.52) If t < T (maximal exposure time, 100 days in this case), compare the exercise and continuation profits: o If Vt0 (t , f1 (t ), f 2 (t )) > Ωt0 (t , f1 (t ), f 2 (t )) , then continue sailing at speed u o If Vt0 (t , f1 (t ), f 2 (t )) = Ωt0 (t , f1 (t ), f 2 (t )) , it is optimal to exercise, so sell the cargo forward and collect Ωt0 (t , f1 (t ), f 2 (t )) 87 • If t = T , sell the cargo forward and collect Ωt0 (T , f1 (T ), f 2 (T )) If the exercise time is t * , the realized profit on this trade is then t* W (t , f1 (t ), f 2 (t )) = ∫ − g (t )dt + Ω(t * , f1 (t * ), f 2 (t * )) * * * (4.53) 0 As can be seen from Figure 39, the realized profit is highly variable – but it stays within the bounds of the 95% confidence interval for W based on the expected risk calculated previously. The standard deviation of W–V calculated over the period when there are excess profits is 6.21 $/bbl, close to the average standard deviation seen in Figure 35. The exposure time, presented in Figure 40, varies widely The annualized Sharpe ratio of the strategy over this period is 1.99 if calculated on the profits in excess of and 5.95 if considered in excess of the risk-free rate. , Realized profits and standard deviation with hedging The significant standard deviation of the realized profits W versus the expected profits V comes from the exposure of the trade to the risk factors f1 and f 2 . Using the hedge ratios computed from the expected profit function V we can simulate what the realized profit is when the profit is delta-hedged with respect to the first or second factor. At time t into the trade, assuming the cargo has not been sold, the value function has deltas δ1 and δ 2 with respect to f1 and f 2 : δk = ∂V (t , f1 (t ), f 2 (t )) ∂f k (4.54) In order to eliminate the risk from factor 1, for example, we take a position −δ1 in the factor f1 . How to achieve this with the available futures contracts is explained in Section 2.9. The impact of this position on the ɶ evolution of the expected portfolio P&L U is ɶ dU = δ1σ 1dW1 + δ 2σ 2 dW2 − δ1df1 (4.55) = δ1 (α1 f1 − µ1 )dt + δ 2σ 2 dW2 Hence the realized P&L at the end of the trade is τ* τ* 0 0 ɶ U (τ , f1 , f 2 ) = V (0, f , f ) + ∫ δ1 (t , f1 (t ), f 2 (t ))(α1 f1 − µ1 )dt + ∫ δ 2 (t , f1 (t ), f 2 (t ))σ 2 dW2 (t ) (4.56) * 0 1 0 2 88 The risk tied to the first factor has therefore been eliminated – but this also reduces the expected profit. The same approach can be applied to the second factor, hedging out tilts. A common practice is to hedge out the parallel shift factor, which is the major risk factor, and keep the exposure to tilts. Statistics 30 Realized profit W ($/bbl) Expected profit V ($/bbl) 95% C. I. for W E[W – V] -0.68 $/bbl Std[W – V] 3.03 $/bbl SR on W 6.48 SR on W – 25 0.06 20 Profit ($/bbl) 15 10 5 0 -5 -10 06/07/07 09/15/07 12/24/07 04/02/08 07/11/08 10/19/08 01/27/09 05/07/09 08/15/09 11/23/09 Figure 41. Expected profit V and realized profit W when hedging the first factor Statistics 30 Realized profit W ($/bbl) Expected profit V ($/bbl) 95% C. I. for W E[W – V] 0.01 $/bbl Std[W – V] 2.42 $/bbl SR on W 9.15 SR on W – 25 1.12 20 Profit ($/bbl) 15 10 5 0 -5 -10 06/07/07 09/15/07 12/24/07 04/02/08 07/11/08 10/19/08 01/27/09 05/07/09 08/15/09 11/23/09 Figure 42. Expected profit V and realized profit W when hedging the first and second factors 89 As can be seen in Figure 41 and Figure 42, the hedging does indeed diminish the risk of the strategy. The historical standard deviation of W − V is • 3.03 $/bbl when hedging the first factor • 2.42 $/bbl when hedging the first and second factor This should be compared to the unhedged standard deviation of 6.21 $/bbl. It is interesting to note that even hedging both factors does not render the strategy riskless, contrary to theory. There are two reasons for this: • The delta-hedging is only daily and not continuous, and high-amplitude movements (jumps) in the factors will not be hedged perfectly • The forward curve does not only move in shifts and tilts, and only those movements have been hedged out 7. Origins of excess profits We have shown that in addition to significant arbitrage profits to be made on arbitraging crude oil between Europe and the United States, following an optimal storage and selling strategy could lead to significant excess profits. It is interesting to understand the origin of these profits in order to understand in what fundamental situations they might appear. We will make a distinction in what follows between • The origin of excess expected profits • The origin of realized profits, i. e. when the trading strategy performs well Origin of excess expected profits We have established in Section 3.5 that the period August 2007 – August 2009 can be decomposed into two periods: August 2007 to October 2008, when the forward curve for crude oil was in backwardation and there were no expected excess profits, and October 2008 to August 2009, when the forward curve was in contango and there could be found excess profits in keeping exposure to the forward curve open. We will concentrate on the second period here. We have already established that the forward curve (net of freight cost) being in contango is a necessary condition for the excess profit to be positive. We can gain more insight into the origins of excess profits by decomposing the excess profit as follows 90 EPtotal = V (0, f1 , f 2 ) − Ω(0, f1 , f 2 ) = EPdrift + EPconvexity + EPearly EPdrift = Ω(T , Ef1 (T ), Ef 2 (T )) − G (T ) − Ω(0, f1 , f 2 ) EPconvexity = Veur − (Ω(T , Ef1 (T ), Ef 2 (T )) − G (T )) (4.57) EPearly = Vearly When considering the initial expected profit, f1 = f 2 = 0 such that Ef1 (T ) = Ef 2 (T ) = 0 and EPdrift = Ω(T , 0, 0) − G (T ) − Ω(0, 0, 0) (4.58) This expected profit will generally be zero for a forward curve in contango. It can, however, be significant for non-zero factor values because of their mean-reverting property. The excess profit from convexity can be written as EPconvexity = E [ Ω(T , f1 (T ), f 2 (T )) ] − Ω(T , Ef1 (T ), Ef 2 (T )) and captures the non-linearity of cargo before the date T. (4.59) . As for the early-exercise premium, it captures the possibility of selling the We present the time series of the excess expected profits and its decomposition in Figure 43. We notice that the major part of the excess profit comes from the convexity, averaging 84% of the total excess profit. The convexity and early-exercise premia are rather regular. 1.5 1 Excess profit ($/bbl) 0.5 0 -0.5 -1 Total Drift Convexity Early -1.5 -2 10/19/08 12/08/08 01/27/09 03/18/09 05/07/09 06/26/09 08/15/09 10/04/09 Figure 43. Decomposition of expected excess profits as a function of time 91 Based on these observations we can conclude that • The existence of an excess profit is conditional on the forward curve net of shipping cost being in contango • When the contango condition is satisfied, the expected excess profits are fairly stable. Trade performance and origin of realized profits We have identified in what situations excess profits are expected. However, in a trading situation it is important to know in what cases the trade will succeed and in which cases it will yield a loss, in order to understand the expected profits and risk manage the position. Profit ($/bbl) 0.2 15 0.1 0.05 10 70 0.04 0.01 40 0 30 -0.01 20 -0.02 10 -0.03 0 U f1 -0.04 -10 f2 0 5 U(t), Ω (t) ($/bbl) 50 f1, f2 60 0.02 f2 0.03 Ω -0.1 0 -0.05 -0.6 -0.4 -0.2 0 f1 0.2 0.4 -0.2 0.6 0 5 10 15 20 Days from trade initiation t 25 -5 30 Figure 44. Evolution of f1, f2, expected P&L U and exercise profit Ω during the trade starting on December 18, 2008. Left, the path of (f1(t), f2(t)) during the trade overlaid on the expected profit V. Right, these functions as a function of days from trade initiation. The delivery of the cargo is specified after 29 days. In Figure 44 we present the evolutions of the factor values and the expected and exercise profits U and during the physical trade initiated on December 18, 2008. In this particular case, the cargo is exercised after 29 days, when the expected profit and exercise profit are seen to converge. The realized profit W at the end of the trade is 13.6 $/bbl. We also present the evolutions of the factor values f1 and f2 on the same figure. As we have already seen, V has the strongest delta with respect to the first factor, and the realized profit is highly correlated with the value of f1 during the trade. When hedging the first factor, however, the realized profits are more correlated with the second factor, as is seen in Figure 45. 92 8 0 6 U(t), Ω (t) ($/bbl) 10 0.1 f1, f2 0.2 Ω -0.1 U f1 4 f2 -0.2 0 5 10 15 20 Days from trade initiation t 25 2 30 Figure 45. Evolution of the factor values and profits during the trade starting on December 18, 2008, when hedging the first factor In order to assess how the trade will perform based on the evolutions of the two factors it is useful to recall the shape of the payoff function as a function of both factors. V(f1,0), tau = 100 days V(0,f2), tau = 100 days 80 1 12 1 0.9 60 0.9 10 V Veur 0.8 0.8 Vearly 40 0.7 8 0.7 Profit ($/bbl) Profit ($/bbl) Ω 0.6 20 0.5 Ω 0 0.4 V early 0.2 -0.6 -0.4 -0.2 0 24 29 34 41 0.3 2 0.2 f* 1 -40 0.1 -0.8 0.5 4 0.4 V V eur 0.3 -20 0.6 6 0.2 0.4 0.6 0.8 0 0.1 -0.06 -0.04 -0.02 0.02 0.04 0.06 49 59 F1 level ($/bbl) 70 84 101 0 +6.04 +5.67 +5.30 +4.93 +4.56 Fwd curve slope ($/bbl/month) +4.19 +3.82 0 f1 0 f2 Figure 46. Cross-section of Ω and V at trade initiation as a function of f1 (left) and f2 (right) As can be seen in Figure 46 the dependence of the trade has the following characteristics with respect to f1 and f2: • It is directional with respect to f1, similar to a forward exposure • It is a volatility trade with respect to f2 : the payoff is convex and has higher payoff for large movements of f2 in either direction. This is closer to a straddle option. 93 Figure 47. Realized profits as a function of realized drifts and volatilities during the trade period, for a trade starting on December 18, 2008 7 7 No hedging Factor 1 hedged Factors 1 & 2 hedged Ref. value 6 6 5 5 No hedging Factor 1 hedged Factors 1 & 2 hedged Ref. value 4 4 W W 3 3 2 2 1 1 0 0 -1 -0.4 -1 -0.3 -0.2 -0.1 0 0.1 Realized µ1 0.2 0.3 -2 0.4 0.2 0.4 0.6 0.8 Realized σ 1 1 1.2 1.4 4.5 4.5 No hedging Factor 2 hedged Factors 1 & 2 hedged Ref. value 4 4 3.5 3 3 W 3.5 W 0 2.5 2.5 2 No hedging Factor 2 hedged Factors 1 & 2 hedged Ref. value 2 1.5 -0.025 1.5 -0.02 -0.015 -0.01 -0.005 0 0.005 Realized µ 0.01 0.015 0.02 0.025 2 0 0.01 0.02 0.03 0.04 0.05 Realized σ 2 0.06 0.07 0.08 This intuition is confirmed by the results in Figure 47. In this figure we present the realized profits as functions of realized drifts and volatilities, imposed in a Monte Carlo simulation, different from the a priori drifts and volatilities used when valuing the floating storage opportunity. We can clearly see the directional nature of the position in f1, with realized profits that are linear in the drift µ1 . These are also increasing in the volatility σ 1 because of the slight convexity of the payoff function. On the other hand, the realized profits are independent of the drift µ 2 of the second factor, but strongly related to its realized volatility. 94 8. General commodity trading problem The problem we have been considering is limited to a single delivery port and a single tanker speed. The only choice that is left to the trader is time of delivery. In general, an oil (or other commodity) cargo that has not yet been sold forward can be rerouted to a different port. The ship can also sail slower in order to save fuel. These optionalities make the cargo more valuable to a trader than what has been calculated previously. A decision model for optimal ship routing should take into account the forward prices at different potential delivery ports and open for the possibility of delaying the choice of delivery port to a later date. For example, a cargo of Bonny Light crude oil loaded in Nigeria could potentially be delivered to Europe or the United States. Instead of choosing a delivery location immediately the oil trader could choose to route the ship northbound in the mid-Atlantic, and waiting to see if the spread evolves. We will formulate the stochastic control equations governing how the ship should be routed to maximize profit. The notations are the same as previously, but we now introduce: • A set of destination ports X kP at which the cargo can be delivered, each with a forward curve Fk (τ ) • The speed of the ship u can be varied within bounds [u1 , u2 ] , usually between 8 and 16 knots • The instantaneous direction of the ship is the unit vector d Exercise profit We define Ω( X , Fk (τ )) to be the profit that can be earned on the cargo if the ship is at location X and the forward curve in port k is Fk (τ ) , by committing to a specific delivery price sometime in the future and sailing to deliver at that time. In effect, the trader gives up the possibility of changing delivery port and time. At exercise, one chooses a delivery port k, a time-to-delivery τ and a sailing speed u. For one choice of these parameters, the profit is τ sail (k , u ) = d ( X , X kP ) / u ω (k, τ , u ) = Fk (τ ) − S0 − Hτ − B ⋅ ( FC (u )τ sail (k , u ) + FCa (τ − τ sail (k , u ))) − BH k (τ ≥ τ sail (k , u )) (4.60) The exercise profit is obtained by maximizing ω ( k, τ , u ) over all possible ports, speeds and times to delivery: Ω( X , Fk (τ )) = max ω (k, τ , u ) k, u τ ≥τ sail ( k, u ) 95 (4.61) Expected profit and optimal route Let us define V ( X , Fk (τ )) as the expected profit from the cargo when the tanker is at the location X and the forward curves are given by Fk (τ ) . Let g ( X , u ) be the daily cost of sailing at speed u when the ship is at location X, i. e. g ( X , u ) = H + B ⋅ FC (u ) if the ship is sailing g ( X , u ) = H + B ⋅ FCa if the ship is at anchor at port k P k (4.62) When the ship is at a location X , the trader has two choices: • either “exercise” and sell the cargo forward, thereby earning the exercise profit Ω( X , Fk (τ )) • or choose to continue speculating during a time dt without exercising. If the ship is at sea he can choose the optimal speed u and direction d and the expected profit is: VC ( X , Fk (τ )) = max E V ( X + udtd , Fk (τ ) + dFk ) − g ( X , u )dt u, d (4.63) If the ship is in port (floating storage), the expected profit is VC ( X lP , Fk (τ )) = E [V ( X , Fk (τ ) + dFk )] − g ( X lP , 0)dt (4.64) This gives the continuation value VC . The forward curves are evolved during the time period dt using the two-factor model. Hence the expected profit at location X given the forward curves Fk (τ ) is V ( X , Fk (τ )) = max [ Ω( X , Fk (τ )), VC ( X , Fk (τ ))] (4.65) Hamilton-Jacobi-Bellman equation When we assume that the underlying factors follow diffusions we can derive a continuous-time equation to evaluate V. This equation is known as the Hamilton-Jacobi-Bellman (HJB) equation for the stochastic control problem, see Morimoto (2010) and Chang (2004). We assume that each of the forward curves Fk (τ ) is governed by a two-factor model, such that k k log Ft k (τ ) = log F0k (t + τ ) +ψ 1k (t, τ ) +ψ 2 (t, τ ) + u1k (τ ) f1k (t ) + u2 (τ ) f 2k (t ) where: 96 (4.66) df jk (t ) = −α k f jk (t )dt + σ k dW jk (t ), j j dψ k (t, τ ) = − j ( j = 1, 2 ) 2 1 kk σ j u j (t + τ ) dt 2 (4.67) For the sake of simplicity we renumber the factors f jk as a sequence ( f i )i =1. M . While the two factors for a single forward curve are uncorrelated, factors for different forward curves will be correlated, such that in general df k df l = σ kσ l ρ kl dt (4.68) Consider a location X, time t and factor values f i and assume that VC > Ω , i. e. we are in the continuation region, s. t. V = VC . We develop V using Ito’s formula: M ∂V ∂V ∂V 1M M ∂ 2V dt + ⋅ d ⋅ dX + ∑ dfi + ∑∑ dfi df j ∂t ∂X 2 i =1 j =1 ∂fi ∂f j i =1 ∂f i (4.69) 2 M MM M ∂V ∂V ∂V 1 ∂V ∂V =V + +u ⋅ d + ∑ ( µi − α i f i ) + ∑∑ ρijσ iσ j σ i dWi dt + ∑ ∂t ∂X ∂fi 2 i =1 j =1 ∂fi ∂f j i =1 i =1 ∂f i V ( X + d ⋅ dX , t + dt , f1 + df1 . f M + df M ) = V + Taking expectations in the definition of VC and simplifying, we finally get the equation: M ∂V ∂V ∂V 1 M M ∂ 2V 0 = max − g ( X , u ) + +u ⋅ d + ∑ ( µi − α i f i ) + ∑∑ ρijσ iσ j u, d ∂t ∂X ∂f i 2 i =1 j =1 ∂f i ∂f j i =1 (4.70) This is a typical example of a Hamilton-Jacobi-Bellman equation. It is valid in the continuation region, i. e. ( f1 . f M ) ∈ S * (t ) . The boundary conditions are given by the smooth pasting condition on the free boundary ∂S * (t ) : V (t , X , f1 . f M ) = Ω(t , X , f1 . f M ) ∀( f1 . f M ) ∈ ∂S * (t , X ), ∇V (t , X , f1 . f M ) = ∇Ω(t , X , f1 . f M ) (4.71) The terminal condition is that at the maximal exposure time T, the cargo should be delivered: V (T , X , f1 . f M ) = Ω(T , X , f1 . f M ) (4.72) Solving the general problem This problem can in principle be solved numerically by dynamic programming or finite differences. However, the potentially large number of state variables can make it challenging to solve using these methods. The preferred method for such a problem would be a least-squares Monte Carlo simulation as presented in Longstaff and Schwartz (2001). This requires work on finding appropriate basis functions to project the solution on. 97 5. CONCLUSIONS 1. Summary of results In Part 2 we have developed a two-factor model and given evidence that it is sufficient for modeling the term structure of volatility and the correlation surface of a number of commodities. We prove that it is easily formulated as a model involving two independent and mean-reverting factors that represent the change in level and slope of the forward curve. We find that the first factor is the dominant factor and the majority of variance of forward prices comes from the first factor. However, other factors cannot be ignored as they will affect portfolios that are weighted differently. We also show that the spot price process implied by this two-factor model is consistent with the Schwartz and Smith (2000) formulation with short-term and long-term shocks driving the spot price. Furthermore, we show that the shapes of forward curves consistent with the two-factor model are exponentials of the factors weighted by their factor loadings. This allows for a simple calibration of forward curves to the market model and an interpretation of the factor values in terms of mean level and initial slope of the curve. The applicability of this model to a number of forward markets, as well as its simple analytical formulation, makes it useful in different valuation settings involving commodity prices. In Part 3 we address the pricing of Asian options written on commodity forwards. We show that by understanding the term structure of volatility correctly, as well as the effect of the averaging on the volatility of the payoff, Asian options can be priced approximately but analytically in a simple way. Comparing our theoretical prices to market prices, we find that it correctly reproduces the term structure of implied volatilities. The understanding of this should increase liquidity in the freight options market. The understanding of volatility and its value also has a profound impact on valuation and operational decisions that involve commodities. In Part 4 we study the floating storage trade involving crude oil and tankers using the two-factor model. This trade can be viewed as the sum of a cross-Atlantic and temporal arbitrage trade – arbitraging crude oil between Europe and the United States and between now and the future – and of a storage trade where the trader can choose the optimal time to release the oil into the market. We show that while the arbitrage window has been open for most of the time during 2007-2009, the storage trade has only existed in the second half of this period. The floating storage opportunity is associated with a forward curve in contango when netted of freight costs. When it is open, there is additional value involved in not selling the cargo immediately and taking advantage of the possibility of higher prices. The framework that we present allows us to evaluate the profits from such a strategy, the decision rules for running the trade, and its exposure to the two risk factors through hedge ratios. The excess value is understood as a combination of the drifts in the factors, of the payoff convexity and of an early exercise premium. 98 2. Suggestions for future research As pertains to the market modeling, an essential improvement that is not performed in the present thesis would be to allow the model to be easily calibrated to market implied volatilities as well as the historical correlation surface. The crude oil market, for example, has a very liquid options market that can be used for such calibration. The market modeling framework presented in this thesis can be applied to a number of problems related to commodity trading. It would be very interesting to see empirical results for the general commodity trading problem presented in Section 4.8, and understand what value is associated with the possibility of switching destinations. It can also be applied to Liquefied Natural Gas cargoes that are currently being rerouted from their long-term contract destinations in the United States to Europe or Asia. Furthermore, this market-based routing problem should be integrated with the optimal weather routing problem developed in Avougleas and Sclavounos (2009). An underlying assumption in our formulation is that routes are deterministic and fuel consumption only depends on speed. In practice, ship routing and fuel consumption depends strongly on weather, and using forecasts and dynamic programming one can determine the optimal route to follow. Integrating this uncertainty with our model would give a much more precise evaluation of the commodity trade, especially when profits come from geographical spreads and not floating storage. However, this general problem involves a large number of state variables and is difficult to solve using dynamic programming. Developing solution methods adapted to such a large-scale problem would greatly enhance its applicability. One promising method, applied for American options, is the Longstaff and Schwartz (2001) least squares Monte Carlo method. This would require finding suitable basis functions on which to project the solution. In this thesis we view the shipping problem from the point of view of a physical oil trader who has the possibility of chartering a ship for one trade, before returning it to the market. Another direction would be to see the problem from the point of view of a shipowner or long-term charterer who can operate the ship continuously on several trades. In that case, the decision taken on one trade, such as storing oil, will have consequences for the next one. In some cases it might be more profitable to sell the oil, return to the loading port and take advantage of a better geographical spread. The same framework can be used, but the problem is of longer-term nature, of years rather than weeks. 99 6. APPENDIX 1. Traded volumes in commodity derivative markets From ICE (2009) and CME (2009): Contract Daily volume (‘000 bbl) ICE Brent Crude Futures Yearly volume (‘000 bbl) 287,355 Brent Crude Options 74,137,750 823 WTI Crude Options 212,341 179,820 WTI Crude Futures 46,393,671 70 18,200 545,351 141,791,260 Crude oil 4,521 1,175,460 Miny WTI 13,369 1,737,970 Brent Financial Futures 1,997 519,220 Dubai Crude oil Calendar 3,665 952,900 WTI Calendar 4,198 1,091,480 546 141,960 NYMEX (Options) NYMEX (Futures) Crude oil physical Brent Calendar Options Brent last day 74 19,240 Crude oil 1mo 2,857 742,820 Crude oil APO 12,713 29,458,520 1,170,661 Total 3,305,380 113,302 Crude oil physical 301,698,172 From Imarex (2009). One lot is 1000 metric tons. Period # trades # lots Dec ‘09 698 14 504 Nov ‘09 1 017 20 817 Oct ‘09 1 083 17 750 Sep ‘09 1 066 13 733 Aug ‘09 711 12 795 Jul ‘09 1 048 14 113 Jun ‘09 1 328 24 766 May ‘09 1 128 16 458 Apr ‘09 1 249 18 703 Mar ‘09 1 362 19 965 Feb ‘09 1 133 15 625 Jan ‘09 1 343 18 020 13 166 207 249 Total 100 2. Spot price process implied by the two-factor model Using the forward curve process and the spot-forward relationship S (t ) = F (t , t ) , we get: dF (t , T ) = σ S e −α (T −t ) dWS + σ L dWL = σ S (t , T )dWS + σ L (t , T )dWL F (t , T ) t log S (t ) = log F (0, t ) − t t t 1 1 2 2 ∫ σ S ( s, t )ds + ∫ σ S (s, t )dWS ( s) − 2 ∫ σ L ( s, t )ds + ∫ σ L ( s, t )dWL (s) 20 0 0 0 Such that: t t ∂ log F (0, t ) 1 2 ∂σ S ( s, t ) ∂σ ( s, t ) d log S (t ) = ds + ∫ S dWS ( s ) − σ S (t , t ) − ∫ σ S ( s, t ) ∂t 2 ∂t ∂t 0 0 t t 12 ∂σ ( s, t ) ∂σ ( s, t ) − σ L (t , t ) − ∫ σ L ( s, t ) L ds + ∫ L dWL ( s ) dt 2 t ∂t 0 0 +σ S (t , t )dWS (t ) + σ L (t , t )dWL (t ) We have: ∂σ S ( s, t ) = −ασ S ( s, t ), ∂t ∂σ L ( s, t ) =0 ∂t Such that ∂σ ( s, t ) ∫ S∂t dWS (s) = −α ∫ σ S (s, t )dWS ( s) 0 0 t t t t t 1 1 2 2 = −α log S (t ) − log F (0, t ) + ∫ σ S ( s, t )ds + ∫ σ L ( s, t )ds − ∫ σ L ( s, t )dW2 ( s ) 20 20 0 Let: µ (t ) = t 1 ∂ log F (0, t ) 1 2 12α 2 2 − σ S (t , t ) − σ L + ∫ (σ S ( s, t ) − σ L )ds + α log F (0, t ) + ασ L (WL (t ) − WL (0)) α ∂t 2 2 20 101 2 1 ∂ log F (0, t ) 1 2 α σS 2 2 µ (t ) = (1 − e −2α t ) − σ L t + α log F (0, t ) + ασ LWL (t ) − (σ S + σ L ) + ∂t 2 2 2α α 2 2 σ 1 ∂ log F (0, t ) σ = + log F (0, t ) − S (1 + e −2α t ) − L (1 + α t ) + σ LWL (t ) 4α 2α ∂t α Then: d log S (t ) = α ( µ (t ) − log S (t ))dt + σ S dWS + σ L dWL d µ (t ) = m(t )dt + σ L dWL where: 1 ∂ 2 log F (0, t ) ∂ log F (0, t ) 1 2 −2α t 2 m(t ) = + + σSe −σL 2 α ∂t ∂t 2 ( ) 3. Principal Components Analysis of the two-factor model We want to find the functions uk that are eigenvectors of the covariance matrix Σ(τ 1,τ 2 ) with associated eigenvalues λk . For this we must choose some arbitrary maximal tenor T, and find eigenvalues λk and eigenvectors uk (τ ) satisfying: τ max ∫ Σ(τ 1,τ 2 )uk (τ 2 )dτ 2 = λk uk (τ 1 ) 0 τ max ∫u 2 k (τ )dτ = 1 0 τ max ∫ u (τ )u (τ )dτ = δ k l kl 0 Given the parametric form of Σ(τ 1,τ 2 ) we find that τ max ∫ 2 (σ S e−ατ1 + ρσ L )(σ S e−ατ 2 + ρσ L ) + (1 − ρ 2 )σ L uk (τ 2 )dτ 2 = λk uk (τ 1 ) 0 Leaving out the k index and developing this equation we find that τ max λu (τ 1 ) = ∫ 0 τ max 2 2 (σ S e −ατ 2 + ρσ Sσ L )u (τ 2 )dτ 2 e−ατ1 + ∫ ( ρσ Sσ L e−ατ 2 + σ L )u (τ 2 )dτ 2 0 102 Thus we see that u (τ ) can be written in the form u (τ ) = Ae −ατ + B where A and B are constants. We replace this expression into the equation to find that: τ max 2 −ατ −ατ τ max −ατ 2 2 2 λσ (τ ) = ∫ (σ S e + ρσ Sσ L )( Ae + B )dτ e + ∫ ( ρσ Sσ L e −ατ 2 + σ L )( Ae−ατ 2 + B )dτ 0 0 Equating the constant and exponential terms we get the matrix eigenvector equation: τ max 2 −ατ −ατ ∫ (σ S e 2 + ρσ Sσ L )e 2 dτ 2 A λ = τ 0 B max 2 ( ρσ Sσ L e−ατ 2 + σ L )e−ατ 2 dτ 2 ∫ 0 τ max 2 (σ S e−ατ 2 + ρσ Sσ L )dτ 2 0 A τ max B −ατ 2 2 ∫ ( ρσ Sσ L e + σ L )dτ 2 0 ∫ This shows that λ is an eigenvalue and ( A, B) an eigenvector of the two-dimensional matrix M. Thus there are only two distinct eigenfunctions u (τ ) and eigenvalues λ - as expected for a twofactor model. 4. Evolution of the constant-maturity forward curve under the two-factor model If we let σ k (t , T ) = σ k uk (T − t ) = σ k ( Ak e −α k (T −t ) + Bk ) , we have: t 2 t 1 log f (t, τ ) = log F (t , t + τ ) = log F (0, t + τ ) + ∑ − ∫ σ k2 ( s, t + τ )ds + ∫ σ k ( s, t + τ )dWk ( s ) 20 k =1 0 Let: sk (t, τ ) = − t t 1 σ k2 ( s, t + τ )ds + ∫ σ k ( s, t + τ )dWk ( s) 2∫ 0 0 Expand the stochastic component (dropping the index k for now): t t 0 t 0 −ατ −α ( t − s ) ∫ σ (s, t + τ )dW (s) = Ae ∫ σ e dW (s) + Bσ ∫ dW ( s) 0 t t 0 0 = ( Ae−ατ + B)σ ∫ e −α (t − s ) dW ( s ) + Bσ ∫ (1 − e−α (t − s ) )dW ( s ) Let: 103 t f (t ) = σ ∫ e −α (t − s ) dW ( s ) 0 t g (t ) = Bσ ∫ (1 − e−α ( t − s ) )dW ( s ) 0 ψ (t, τ ) = − t 1 σ 2 ( s, t + τ )ds 2∫ 0 Then s (t, τ ) = ψ (t, τ ) + g (t ) + u (τ ) f (t ) Differentiate this df (t ) = −α f (t )dt + σ dW (t ) dg (t ) = B(σ dW + α f (t )dt − dW ) = Bα f (t )dt t 1 ∂σ dψ (t, τ ) = − σ 2 (t , t + τ ) − ∫ σ ( s, t + τ ) ( s, t + τ )ds dt = µ (t, τ )dt ∂T 0 2 We recognize that f (t ) is an Ornstein-Uhlenbeck process mean-reverting to 0, g (t ) is an integral of f and ψ (t, τ ) is a deterministic drift. We can calculate µ k (t, τ ) explicitly 1 2 t µk (t, τ ) = − σ k2 (t , t + τ ) − ∫ σ ( s, t + τ ) 0 ∂σ ( s, t + τ )ds ∂T t 1 = − σ k2 ( Ak e−α kτ + Bk ) 2 + σ k2α k ∫ ( Ak e−α k ( t +τ − s ) + Bk )Ak e −α k ( t +τ − s ) ds 2 0 1 1 = − σ k2 ( Ak e−α kτ + Bk ) 2 + σ k2 Ak2e −2α kτ (1 − e−2α k t ) + σ k2 Ak Bk e −α kτ (1 − e−α k t ) 2 2 1 1 = − σ k2 Ak2 e−2α k ( t +τ ) − σ k2 Ak Bk e−α k ( t +τ ) − σ k2 Bk2 2 2 1 1 µk (t, τ ) = − σ k2 ( Ak e−α k ( t +τ ) + Bk ) 2 = − σ k2uk (t + τ ) 2 2 2 Thus the constant-maturity futures price can be written as: log f (t, τ ) = log F (0, t + τ ) + ψ 1 (t, τ ) + ψ 2 (t, τ ) + g1 (t ) + g 2 (t ) + u1 (τ ) f1 (t ) + u2 (τ ) f 2 (t ) where: 104 df k (t ) = −α k f k (t )dt + σ k dWk (t ) dg k (t ) = Bkα k f k (t )dt 1 dψ k (t, τ ) = − σ k2uk2 (t + τ )dt 2 5. Impact of a third factor on the constant-maturity forward curve We have, as in Appen dix 4, that t 3 t 1 log f (t, τ ) = log F (t , t + τ ) = log F (0, t + τ ) + ∑ − ∫ σ k2 ( s, t + τ )ds + ∫ σ k ( s, t + τ )dWk ( s ) 20 k =1 0 Let: t t 1 sk (t, τ ) = − ∫ σ k2 ( s, t + τ )ds + ∫ σ k ( s, t + τ )dWk ( s ) 20 0 We consider only the third factor and will assume k = 3 in what follows. Let us consider first the stochastic part: t t 0 0 ∫ σ (s, t + τ )dW (s) = σ ∫ ( Ae t −2α ( t +τ − s ) = u (τ ) ∫ σ e −2α ( t − s ) + Be−α ( t +τ − s ) + C )dW ( s ) dW ( s ) + Be −ατ 0 t ∫σ e −α ( t − s ) (1 − e −α (t − s ) )dW ( s ) 0 t + C ∫ σ (1 − e−2α (t − s ) )dW ( s ) 0 Let: t t t 0 0 0 f (t ) = ∫ σ e−2α (t − s ) dW ( s ), g (t ) = ∫ σ e −α (t − s ) (1 − e −α (t − s ) )dW ( s ), h(t ) = ∫ σ (1 − e −2α (t − s ) )dW ( s ) Then df (t ) = −2α f (t ) + σ dW (t ) (Ornstein-Uhlenbeck process) dg (t ) = α ( f (t ) − g (t ))dt g (t ) = α ∫ e−α ( t − s ) f ( s )ds t 0 dh(t ) = 2α f (t )dt t h(t ) = 2α ∫ f ( s )ds 0 105 The process f (t ) is an Ornstein-Uhlenbeck process mean-reverting to zero with mean-reversion speed 2α and volatility σ . The processes g (t ) and h(t ) are stochastic drifts – integrals of f (t ) with different weights. 6. Black volatilities of the Average price contract In-settlement We consider a date TM ≤ t < TM +1 . 2 2 1 − e −α (T − s ) T −s 2 T −s 2 σ ( s, T )ds = ∫ σ S + ρ ' σ L + (1 − ρ ) ' σ L ds ' ∫ α cM cM cM t t T T 2 A = T −t ∫ 0 2 σS 2 ρσ Sσ L s2 2 −α s 2 −α s (1 − e ) s + ' 2 σ L ds 2 ' 2 (1 − e ) + ' cM α cM 2 α cM Let us calculate each of the terms separately: T −t ∫α 0 T −t ∫ 0 2 σS 2'2 M c (1 − e −α s ) 2 ds = 2 σS 2 1 − e −2α (T −t ) −α ( T − t ) + T − t − 1− e ' α 2 cM 2 α 2α 2 ρσ Sσ L 2 ρσ Sσ L (1 − e−α s ) sds = '2 ' α cM α cM 2 = ( ) s 2 T − t e −α s T −t e −α s T −t − s + 2 2 0 −α 0 α 0 2 ρσ Sσ L (T − t ) 2 (T − t )e −α (T −t ) 1 + − 2 1 − e −α (T −t ) '2 α cM 2 α α ( T −t ∫ 0 s2 2 1 (T − t )3 2 σ L ds = σL ' ' cM 2 3 cM 2 Such that the square of the Black volatility is given by: 106 ) σ Black (t , T ) 2 = = T 1 2 ∫ σ A (s, T )ds T −t t 2 σS 2 1 − e−α (T −t ) 1 1 − e −2α (T −t ) + 1 − +⋯ ' α 2 cM 2 α T − t 2α T −t 2 ρσ Sσ L T − t e −α (T −t ) 1 1 − e−α (T −t ) + −2 +⋯ ' α cM 2 2 α α T −t 1 (T − t ) 2 2 σL ' 3 cM 2 Let us consider the case when α c ≪ 1 and simplify this expression σ Black (t , T ) 2 ≈ σ Black (t , T ) ≈ 2 2 ρσ Sσ L 1 σS 1 σ2 (T − t ) 2 + (T − t )2 + ' L2 (T − t ) 2 ' ' 3 cM 2 3cM 2 3 cM 1 T −t 2 2 σ S + 2 ρσ Sσ L + σ L ' 3 cM ' and cM = TN − TM ≈ T − t such that σ Black (t , T ) ≈ 1 2 2 σ S + 2 ρσ Sσ L + σ L 3 Pre-settlement T1 T ∫σ 2 A 2 2 ( s, T )ds = ∫ σ A ( s, T )ds + (T − T1 )σ Black (T1 , T ) t t The second term is known from the calculations above. Let us calculate the first term. 2 αc −1 −α ( T − s ) e 2 σ ( s, T )ds = ∫ σ S e + ρσ L + (1 − ρ 2 )σ L ds ∫ αc t t T1 T1 2 A 2 T1 2 −2α (T − s ) S = ∫σ e t 1 eα c − 1 eα c − 1 −α (T − s ) 2 e ds + (T1 − t )σ L ds + ∫ 2 ρσ Sσ L αc αc t T 2 eα c − 1 e−2α (T −T1 ) − e−2α (T −t ) eα c − 1 e−α (T −T1 ) − e−α (T −t ) 2 =σ + 2 ρσ Sσ L + (T1 − t )σ L αc α 2α αc 2 S If we assume that α c ≪ 1 , and noticing that T − T1 = c 107 e−2α c − e−2α (T −t ) e −α c − e −α ( T − t ) c2 + 2 ρσ Sσ L + 1 − σ L 2α (T − t ) α (T − t ) T −t c1 2 2 + (σ S + 2 ρσ Sσ L + σ L ) T −t 3 c 1 − e−2α (T1 −t ) 1 c 1 − e −α (T1 −t ) 1 2c 2 2 σ Black (t , T ) 2 ≈ σ S + + 2 ρσ Sσ L + + σ L 1 − αc T − t 2α c 3 T −t 3 3 T −t 2 σ Black (t , T ) 2 ≈ σ S We check that when t = T1 (i. e the contract enters settlement): 1 3 2 σ Black (T1 , T ) 2 ≈ σ S + 2 ρσ Sσ L 1 2 + σL 3 3 and when T − t → ∞ , σ Black (t , T ) → σ 2 7. Semi-analytical solution to the optimal stopping problem We begin by presenting the analysis in the simple case of one factor. The continuation region is then given by S * (t ) = (−∞, f * (t )) . The equation satisfied by the value function V is ∂V ∂V 1 2 ∂ 2V + (µ − α f ) +σ − g (t ), ∂t ∂f 2 ∂f 2 f < f * (t ) V (t , f ) = Ω(t , f ), f ≥ f * (t ) V (T , f ) = Ω(T , f ) Let LV = (µ − α f ) ∂V 1 2 ∂ 2V +σ − g (t ) ∂f 2 ∂f 2 such that ∂V + LV = ψ (t , f ), ∂t 0 ψ (t , f ) = ∂Ω + LΩ ∂t The transition density for the Ornstein-Uhlenbeck process is 108 f < f * (t ) f ≥ f * (t ) p ( f ',τ , f , t ) = 1 1/2 1 − e−2α (τ −t ) 2πσ 2 2α 2 −α (τ −t ) µ −α (τ − t ) f '− fe + (1 − e ) α 1 exp − 2 1 − e−2α (τ −t ) σ2 2α This transition density satisfies the equation ∂p ∂p 1 ∂2 p + (µ − α f ) + σ 2 2 = 0 ∂t ∂f 2 ∂f The value function can then be written as ∞ V (t , f ) = T∞ ∫ −∞ p ( f ', T ; f , t )(Ω(T , f ') − G (t , T ))df ' + ∫ ∫ p( f ', s; f , t )(−ψ (s, f '))df ' ds t −∞ = Veur (t , f ) + Vearly (t , f ) where G (t , T ) is the cost of shipping between times t and T : T G (t , T ) = ∫ g ( s )ds t Let us verify this result by differentiating the above formula. Consider first the European value Veur (t , f ) : ∞ ∞ ∂Veur ∂p ∂G =∫ ( f ', T ; f , t )(Ω(T , f ') − G (t , T ))df ' − ∫ p ( f ', T ; f , t ) df ' ∂t ∂t ∂t −∞ −∞ = −( µ − α f ) ∂Veur 1 2 ∂ 2Veur −σ + g (t ) 2 ∂f ∂f 2 and the early-exercise premium: ∂Vearly ∂t = +∞ ∫ −∞ T +∞ p( f ', t ; f , t )ψ (t , f ')df ' + ∫ t T +∞ ∂p ( f ', s; f , t ) (−ψ ( s, f '))df ' ds ∂t −∞ ∫ ∂p 1 ∂2 p −( µ − α f ) − σ 2 2 (−ψ ( s, f '))df ' ds ∫ ∂f 2 ∂f t −∞ 2 ∂Vearly 1 2 ∂ Vearly = −( µ − α f ) −σ +ψ (t , f ) 2 ∂f ∂f 2 = ψ (t , f ) + ∫ Furthermore, Veur and Vearly satisfy the terminal conditions 109 Veur (T , f ) = ∞ ∫ p( f ', T ; f , T )(Ω(T , f ') − G (T , T ))df ' = Ω(T , f ) −∞ T∞ Vearly (T , f ) = ∫ ∫ p( f ', s; f , T )(−ψ ( s, f '))df ' ds = 0 T −∞ such that V = Veur + Vearly solves the equation. The early exercise premium can be written in terms of the stopping boundary as: T Vearly (t , f ) = ∫ ∞ ∫ p( f ', s; f , t )(−ψ ( s, f '))df ' ds t f * (s) This formulation gives a closed form expression of V. However, it involves the values of f * ( s ) for t ≤ s ≤ T . These are determined by the continuity condition Ω(t , f * (t )) = V (t , f * (t )) = ∞ ∫ −∞ ∞ T p( f ', T ; f * (t ), t )(Ω(T , f ') − G (t , T )) df ' + ∫ p( f ', s, f * (t ), t )(−ψ ( s, f '))df ' ds ∫ * t f (s) The value function at maturity t and the stopping boundary f * (t ) can be determined recursively as follows: discretize the dates as t0 = 0, t1 . t N = T . If f * (t N ). f * (tk +1 ) have been calculated, let F ( f k* ) = ∞ ∫ p ( f ', T , f k* , t )(Ω(T , f ') − G (t , T ))df ' + −∞ N −1 ∑ ∆t j = k +1 ∞ ∫ p ( f ', t j , f k* , tk )(−ψ (t j , f '))df ' f * (t j ) ∞ + ∆t ∫ p ( f ', tk ; f k* , tk )(−ψ (tk , f '))df ' f k* Finding the stopping boundary f * (tk ) at time tk involves finding, numerically, f * (tk ) = min f k* , F ( f k* ) = Ω(tk , f k* ) Once this stopping boundary has been located the value function can be calculated for all f using V (t , f ) = ∞ ∫ N −1 ∞ j =k f * (t j ) p ( f ', T , f, τ )(Ω(T , f ') − G (t , T ))df ' + ∑ ∆t −∞ ∫ p ( f ', t j , f , tk )(−ψ (t j , f '))df ' We can extend this analysis to two factors, by using the fact that they are independent. The transition density function for the joint Ornstein-Uhlenbeck process is 110 p ( f ',τ ; f , t ) = p1 ( f1',τ ; f1 , t ) p2 ( f 2',τ ; f 2 , t ) where p1 and p2 are the transition densities for the one-dimensional Ornstein-Uhlenbeck processes. This two-dimensional transition density function solves the partial differential equation ∂p ∂p 1 2 ∂ 2 p ∂p 1 2 ∂ 2 p + ( µ1 − α1 f1 ) + σ1 + ( µ2 − α 2 f 2 ) + σ2 2 = 0 ∂t ∂f1 2 ∂f12 ∂f 2 2 ∂f 2 The equation for the value function is ∂V + LV = ψ (t , f ) ∂t 0 ψ (t , f ) = ∂Ω + LΩ ∂t f ∈ S * (t ) f ∉ S * (t ) where L V = ( µ1 − α1 f1 ) ∂V ∂V 1 2 ∂ 2V 1 2 ∂ 2V + ( µ2 − α 2 f 2 ) + σ1 + σ 2 2 − g (t ) ∂f1 ∂f 2 2 ∂f12 2 ∂f 2 such that the value function can be written as T V (t , f ) = ∫ p( f ', T ; f , t )(Ω(T , f ') − G (t , T ))df ' + ∫ ∫ ℝ2 p ( f ', s; f , t )(−ψ ( s, f '))df ' ds t ℝ2 \ S * ( s ) = Veur (t , f ) + Vearly (t , f ) In the two-dimensional case the continuation region S * (t ) is defined as S * (t ) = The boundary ∂S * (t ) of this domain has to be determined for each date t . We write it as a function of the second factor ∂S * (t ) = ( f1* (t , f 2 ), f 2 ), f 2 ∈ ℝ such that the equation to be solved by f1* (t , f 2 ) is 111 V (t , f1* (t , f 2 ), f 2 ) = ∫ p( f f , T ; f1* (t , f 2 ), f 2 , t )(Ω(T , f ') − G (t , T ))df ' '' 12 ℝ2 T +∞ +∫ +∞ p1 ( f1' , f 2' , s; f1* (t , f 2 ), f 2 , t )(−ψ ( s, f '))df ' ds ∫∫ t −∞ f1* ( s , f 2 ) = Ω(t , f1* (t , f 2 ), f 2 ) To find the function f1* (t , f 2 ) we proceed recursively as in the one-factor case. Having determined f1* (t j , f 2 ) for j > k , we calculate F ( f1* , f 2 ) = ∫ p( f , f , T ; f ' 1 ' 2 * 1 , f 2 , tk )(Ω(T , f ') − G (tk , T ))df1'df 2' ℝ2 + N −1 ∑ j = k +1 +∞ ∆t ∫ ∞ p ( f1' , f 2' , t j ; f1* , f 2 , tk )(−ψ (t j , f '))df1'df 2' ∫ −∞ f1* ( t j , f 2' ) + ∆τ (−ψ (t k , f1* , f 2 )) and we find the exercise boundary by varying f1* : f1* (tk , f 2 ) = min f1* , F ( f1* , f 2 ) = Ω(tk , f1* , f 2 ) Once this exercise boundary has been located the value function can be calculated for all f using V (tk , f1 , f 2 ) = ∫ N −1 +∞ j =k −∞ f1* ( t j , f 2 ) p ( f ', T ; f , tk )(Ω(T , f ') − G (tk , T ))df ' + ∑ ∆τ ℝ2 112 ∞ ∫∫ p ( f ', t j ; f , tk )( −ψ (t j , f ')) df ' 8. Routes, cargoes and ships used in the floating storage trades Crude oil Sullom Voe – LOOP Route Sullom Voe – LOOP Ship1 Distance d 4535 Nm Type Very Large Crude Carrier (VLCC) Cargo Brent Cargo size 270 000 mt Barrel factor 7.578 bbl/mt DWT 300 000 mt Loading port Sullom Voe Speed u 15 knots Loading price S0 Dated Brent 10-21 days Fuel consumption sailing: FC(u) anchor: FCa 87.5 mt/day (laden) 74 mt/day (ballast) 85 mt/day (pumping) 15 mt/day (anchor) Loading delay τload 15 days Timecharter price H IFO price B Fuel Oil 3.5% CIF NWE (Platts) Delivery port VLCC average timecharter equivalent (Baltic Exchange) LOOP Delivery price F(t, τ) LLS forward curve Heating oil ARA – NYH Route ARA – NYH Ship Distance d 3383 Nm Type Very Large Crude Carrier (VLCC) Cargo No. 2 fuel oil Cargo size 270 000 mt Barrel factor 312.63 gal/mt DWT 300 000 mt Loading port Amsterdam-Rotterdam-Antwerp Speed u 15 knots Loading price S0 ICE Gasoil front month price Fuel consumption sailing: FC(u) anchor: FCa 87.5 mt/day (laden) 74 mt/day (ballast) 85 mt/day (pumping) 15 mt/day (anchor) Loading delay τload 15 days Timecharter price H IFO price B Fuel Oil 3.5% CIF NWE (Platts) Delivery port VLCC average timecharter equivalent (Baltic Exchange) New York Harbor Delivery price F(t, τ ) Nymex heating oil forward curve 1 Corresponds to the modern double-hull VLCC from Clarksons (2009) 113 7. REFERENCES Albanese, C. & Campolieti, G. (2006). Advanced Derivatives Pricing and Risk Management. Elsevier Academic Press. Avougleas, K. & Sclavounos, P. D. (2009). Fuel efficient ship routing. Massachusetts Institute of Technology, Working paper. Baker, M. P., Mayfield, E. S., & Parsons, J. E. (1998). Alternative models of uncertain commodity prices for use with modern asset pricing methods. The Energy Journal, 19, 1, 115-148 Basilevsky, A. (1994). Statistical Factor Analysis and Related Methods. John Wiley & amp; 자제. Black, F. (1976). The pricing of commodity contracts. Journal of Financial Economics, 3, 167-179. Brennan, M. J. & Schwartz, E. S. (1985). Evaluating Natural Resource Investments. The Journal of Business, Vol. 58. No. 2 (Apr. 1985), pp. 135-157 Carr, P., Ewald, C.-O., & Xiao, Y. (2008). On the qualitative effect of volatility and duration on prices of Asian options. Finance Research Letters, 5, pp. 162-171. Central Intelligence Agency (2010). Country Comparison: oil production. CIA World Factbook, cia. gov/library/publications/the-world-factbook/. Accessed May 7, 2010. Chang, F.-R. (2004). Stochastic Optimization in Continuous Time. Cambridge University Press. Clarksons (2009). Sources & Methods for the Shipping Intelligence Weekly, May 2009. Clewlow, L. & Strickland, C. (1998).Implementing Derivative Models. John Wiley & amp; 자제. Clewlow, L. & Strickland, C. (2000). Energy Derivatives. Pricing and Risk Management. Lacima Publications. CME Group (2009). NYMEX/COMEX Exchange ADV Report – Monthly, December 2009. Dewynne, J. N. & amp; Wilmott, P. (1995). A Note on Average Rate Options with Discrete Sampling. SIAM Journal on Applied Mathematics, Vol. 55, No. 1 (Feb. 1995), pp. 267-276 Dixit, A. K., & Pindyck, R. S. (1994). Investment Under Uncertainty. Princeton University Press. Geman, H., & Yor, M. (1992) Quelques relations entre processus de bessel, options asiatiques, et fonctions confluences hypergeometriques. Comptes Rendus De l’Académie Des Sciences De Paris, I, 471-474. Gibson, R. & Schwartz, E. S. (1990). Stochastic Convenience Yield and the Pricing of Oil Contingent Claims. The Journal of Finance, Vol. 53, No. 3 (Jul. 1990), pp. 959-976. Haug, E. G. (2006). The Complete Guide to Option Pricing Formulas. McGraw-Hill, 2nd edition. 114 Haykov, J. M. (1993). A better control variate for pricing standard Asian options. Journal of Financial Engineering, Vol. 2, No. 3, pp. 207-216. Heath, D., Jarrow, R. A. & amp; Morton, A. (1992) Bond Pricing and the Term Structure of Interest Rates: A New Methodology for Contingent Claims Valuation. Econometrica, Vol. 60, No. 1, pp. 77-105. InterContinental Exchange (2009). ICE Report Center, MTD/YTD Volume for 2009. theice/marketdata/reports/ReportCenter. shtml. Accessed May 7, 2010. International Maritime Exchange (2010). Tanker FFA Trading Statistics. exchange. imarex/information-services/volume-statistics/tanker-ffa/. Accessed May 7, 2010. Joy, C., Boyle, P. P. & amp; Tan, K. S. (1996). Quasi-Monte Carlo Methods in Numerical Finance. Management Science, Vol. 42, No. 6 (June 1996). pp. 926-938. Kemna, A. G.Z & Vorst, A. C.F. (1989). A pricing method for options based on average asset values. Journal of Banking and Finance, Vol. 14, pp. 113-129. Koekebakker, S., & Ollmar, F. (2005). Clarkson Securities Freight Option Model – Technical Document. clarksonsecurities/files/pdf/clarkson_technical_doc. pdf. Koekebakker, S., Ådland, R., & Sødal, S. (2007). Pricing freight rate options. Transportation Research Part E, 43, 535-548. Levy, E. (1992). Pricing european average rate options. Journal of International Money and Finance, 14, 474-491. Longstaff, F. A., Santa-Clara, P. & Schwartz, E. S. (2001) Throwing away a billion dollars: The Cost of Suboptimal Exercise Strategies in the Swaptions Market. Journal of Financial Economics, Vol. 62, No. 1 (Oct. 2001), pp. 39-66. Longstaff, F. A. & Schwartz, E. S. (2001) Valuing American Options by Simulation: A Simple Least-Squares Approach. The Review of Financial Studies, Vol. 14, No. 1, pp. 113-147 Morimoto, H. (2010). Stochastic control and mathematical modeling. Applications in Economics. Cambridge University Press Miltersen, K. R. & amp; Schwartz, E. S. (1998). Pricing of Options on Commodity Futures with Stochastic Term Structures of Convenience Yields and Interest Rates. The Journal of Financial and Quantitative Analysis, Vol. 33, No. 1 (Mar., 1998), pp. 33- 59 Musiela, M. & Rutkowski, M. (2008). Martingale Methods in Financial Modeling. Springer. Paddock, J. L., Siegel, D. R. and Smith, J. L. (1988), Option Valuation of Claims on Real Assets: The Case of Offshore Petroleum Leases. The Quarterly Journal of Economics, Vol. 103, No. 3 (Aug. 1988), pp 479508. Pilipovic, D. (2007). Energy Risk. Valuing and Managing Energy Derivatives. McGraw Hill. 115 Rebonato, R. (2002). Modern pricing of interest rate derivatives: The LIBOR market model and beyond. Princeton University Press. Rogers, L. C.G & Shi, Z. (1995). The Value of an Asian Option. Journal of Applied Probability, Vol. 32, No. 4 (Dec. 1995), pp. 1077-1088. Samuelson, P. A. (1965). Proof that properly anticipated prices fluctuate randomly. Industrial Management Review, Vol. 6, No. 2 (Spring 1965), pp. 41-49 Schwartz, E., & Smith, J. E. (2000). Short-term variations and long-term dynamics in commodity prices. Management Science, 46, 7, 893-911. Sclavounos, P. D., & Ellefsen, P. E. (2009). Multi-factor model of correlated commodity forward curves for crude oil and shipping markets. 매사추세츠 공과 대학. Working Paper. Center for Energy and Environmental Policy Research (CEEPR) Stopford, M. (2009). Maritime Economics. Routledge, 3rd edition. Turnbull, S. M. & amp; Wakeman, L. M. (1991). A Quick Algorithm for Pricing European Average Options. The Journal of Financial and Quantitative Analysis, Vol. 26, No. 3 (Sep. 1991), pp. 377 -389. Yergin, D. (2008). The Prize: The Epic Quest for Oil, Money & power. Free Press, New edition. Ådland, R. (2003). The stochastic behavior of spot freight rates and the risk premium in bulk shipping. PhD Thesis, Massachusetts Institute of Technology. 116 .

TERM Fall '12 PROFESSOR Madu TAGS . crude oil.

문서 세부 정보를 편집하려면 누릅니다.

친구와이 링크 공유 :

Most Popular Documents for MSC GBF.

Questions European Business School London MSC GBF - Fall 2012 Important points to remember while estimating the cash flows of a project are: I) Onl.

Additional_Slides session 1 European Business School London MSC GBF - Fall 2012 Balance Sheet ASSETS LIABILITIES & EQUITY Fixed Assets (non current) Plant & Property.

Additional_Slides session 1.

Business Case lesson 3 European Business School London MSC GBF - Fall 2012 ESCP Europe Masters in Management Business law module, seminar 3 Case.

Business Case lesson 3.

LLP members sentence about workers European Business School London MSC GBF - Fall 2012 LLP members not whistling in the dark Dentons Michael Bronstein, Ryan Carthew , Gilla.

LLP members sentence about workers.

An Empirical Study On Commodity Linked Bonds - Pricing With Monte-Carlo Simulation And Tree Approxim European Business School London MSC GBF - Fall 2012 AN EMPIRICAL STUDY ON COMMODITY LINKED BONDS: PRICING WITH MONTE-CARLO SIMULATION AND.

An Empirical Study On Commodity Linked Bonds - Pricing With Monte-Carlo Simulation And Tree Approxim.

Lecture 7 European Business School London MSC GBF - Fall 2012 Corporate Finance Nov-16 Corporate Finance Dr. Kostas Andriosopoulos [ protected]

최근에 본.

Commodity Market Modeling And Physical Trading Strategies.

이동에 대해 알아보십시오.

기타 관련 자료.

Proof This follows by the one dimensional Itˆo Formula for jump processes European Business School London MSC GBF - Fall 2012 D EPT. OF M ATH ./CMA P URE M ATHEMATICS ISSN 08062439 U NIVERSITY OF O SLO N O 20 O.

A Multivariate Non-Gaussian Stochastic Volatility Model With Leverage For Energy Markets.

From the Appendix the partial derivatives of Π 0 are at most log linearly European Business School London MSC GBF - Fall 2012 Asymptotic Pricing of Commodity Derivatives using Stochastic Volatility Spot Models.

Asymptotic Pricing Of Commodity Derivatives Using Stochastic Volatility Spot Models.

Left The market future prices red dashed line and the forward prices as European Business School London MSC GBF - Fall 2012 A Hybrid Commodity and Interest Rate Market Model K. F. Pilz and E. Schlgl University.

A Hybrid Commodity And Interest Rate Market Model.

cases of knowing the parameters estimating them with maximum likelihood and European Business School London MSC GBF - Fall 2012 Ambiguity, Illiquidity, and Hedge Funds: An Analysis of Recent Developments and Curre.

Ambiguity, Illiquidity, And Hedge Funds - An Analysis Of Recent Developments And Current Research To.

A Closer Look At The Relation Between Garch And Stochaistic Autoregresive Volatility European Business School London MSC GBF - Fall 2012 A Closer Look at the Relation between GARCH and Stochastic Autoregressive Volatility.

A Closer Look At The Relation Between Garch And Stochaistic Autoregresive Volatility.

A New Approach To Markov-Switching Garch Models European Business School London MSC GBF - Fall 2012 A New Approach to Markov-Switching GARCH Models Markus Haasa Stefan Mittnika b c Mar.

A New Approach To Markov-Switching Garch Models.

코스 영웅을 얻으십시오.

적법한.

우리와 접촉.

저작권 © 2017. Course Hero, Inc. 개인 정보 보호 약관.

코스 히어로 (Course Hero)는 대학이나 대학교에서 후원하거나 승인하지 않습니다.

How to leverage market contango and backwardation.

By Jean Folger | February 1, 2013.

While the word contango may sound mysterious, it is used to describe a fairly normal pricing situation in futures. 선물 계약의 선물 가격이 미래의 현물 가격보다 높을 때 시장은 콘탱고에 있다고합니다. 본질적으로 contango의 반대 인 정상 backwardation은 선물 계약의 선물 가격이 미래의 현물 가격보다 낮을 때 발생합니다. Because contango and backwardation are known states in the market, traders can employ strategies that attempt to exploit them.

Contango and backwardation are frequently seen in commodity markets where certain factors prompt the price discrepancy between expected future spot prices and the price of futures contracts. Here, we will introduce the two market states of contango and normal backwardation, explain why they happen and how they affect futures traders.

A brief review of exactly what a futures contract is can be helpful when trying to understand contango and backwardation. A futures contract is a legally binding agreement to buy or sell a specified financial instrument or physical commodity at a predetermined price in the future. The buyer of a futures contract (who is long the contract) expects price to increase, while the seller (who is short the contract) anticipates that price will decrease. 선물 계약은 미래에 무언가를해야 할 의무입니다.

일부 선물 계약은 현금으로 결제되는 반면, 일부 선물 계약은 현금으로 결제됩니다. 많은 상품 선물이 물리적으로 배달됩니다. however, market participants can hedge or speculate in the contract without ever taking physical delivery by entering an offsetting position before the contract expiration date.

Commodities are bought and sold on two separate but associated markets: The cash market, which involves the buying and selling of physical commodities, and the futures market, which involves the buying and selling of a future obligation. In a cash market, transactions are settled on the spot (thus the name “spot” market), meaning that the exchange between the buyer and seller takes place in the present. An instrument’s spot price is its market price, or the price at which it could be bought or sold today (often the futures nearest expiration is considered the spot futures).

In a futures market, on the other hand, any exchange between the buyer and seller takes place at some predetermined time in the future. The various futures contracts (for example, the February 2013, March 2013 and April 2013 light sweet crude oil futures contracts) state the price that will be paid and the date of delivery. 대부분의 계약은 만료되기 전에 직위를 상쇄하므로 계약이 거의 이루어지지 않습니다.