Generalized Multi-Factor Commodity Spot Price Modeling through Dynamic Cournot Resource Extraction Models

Transcription

1 Generalized Multi-Factor Commodity Spot Price Modeling through Dynamic Cournot Resource Extraction Models Bilkan Erkmen (joint work with Michael Coulon) Workshop on Stochastic Games, Equilibrium, and Applications to Energy & Commodities Markets Fields Institute, University of Toronto, Toronto, Canada August 28, 2013

2 Background

3 Background Commodity contracts exhibit certain empirical regularities: mean-reversion and convenience yield in spot prices, backwardation, contango and decreasing volatility term structure (Samuelson Effect) in futures prices

4 Background Commodity contracts exhibit certain empirical regularities: mean-reversion and convenience yield in spot prices, backwardation, contango and decreasing volatility term structure (Samuelson Effect) in futures prices There is a significant body of literature on commodity price modeling that attempts to capture these empirical observations

5 Background Commodity contracts exhibit certain empirical regularities: mean-reversion and convenience yield in spot prices, backwardation, contango and decreasing volatility term structure (Samuelson Effect) in futures prices There is a significant body of literature on commodity price modeling that attempts to capture these empirical observations Modeling efforts are broadly categorized into two groups:

6 Background Commodity contracts exhibit certain empirical regularities: mean-reversion and convenience yield in spot prices, backwardation, contango and decreasing volatility term structure (Samuelson Effect) in futures prices There is a significant body of literature on commodity price modeling that attempts to capture these empirical observations Modeling efforts are broadly categorized into two groups: 1. Reduced-form models that characterize the spot price as the solution to an SDE: Schwartz 1997, Schwartz & Smith 2000, Hilliard & Reis 1998

7 Background Commodity contracts exhibit certain empirical regularities: mean-reversion and convenience yield in spot prices, backwardation, contango and decreasing volatility term structure (Samuelson Effect) in futures prices There is a significant body of literature on commodity price modeling that attempts to capture these empirical observations Modeling efforts are broadly categorized into two groups: 1. Reduced-form models that characterize the spot price as the solution to an SDE: Schwartz 1997, Schwartz & Smith 2000, Hilliard & Reis Structural models that explicitly capture the supply - demand intraction, market mode (monopoly, oligopoly or competitive) and other economic dynamics: Sundaresan 1984, Reinganum & Stokey 1985, Dockner et al. 2001, Sircar et al. 2009

8 Research Motivation While reduced-form and structural models are somehow connected (economic intuition underlying both types of models is similar), there does not appear to be an attempt to formally unite them. The research motivation for this work is this absence of investigation into the precise relationship between reduced form and structural models of commodity prices.

9 Literature Review (1/2) Schwartz 1997 reviews multi-factor commodity spot price models:

10 Literature Review (1/2) Schwartz 1997 reviews multi-factor commodity spot price models: One-factor exponential Ornstein-Uhlenbeck mean-reverting model for the commodity spot price: ds t = κ (µ ln (S t)) S tdt + σs tdz t

11 Literature Review (1/2) Schwartz 1997 reviews multi-factor commodity spot price models: One-factor exponential Ornstein-Uhlenbeck mean-reverting model for the commodity spot price: ds t = κ (µ ln (S t)) S tdt + σs tdz t Two-factor model w/ stochastic convenience yield: ds t = (µ δ t) S tdt + σ 1 S tdzt 1 dδ t = κ (α δ t) dt + σ 2 dzt 2 where dzt 1dZ t 2 = ρdt.

12 Literature Review (1/2) Schwartz 1997 reviews multi-factor commodity spot price models: One-factor exponential Ornstein-Uhlenbeck mean-reverting model for the commodity spot price: ds t = κ (µ ln (S t)) S tdt + σs tdz t Two-factor model w/ stochastic convenience yield: ds t = (µ δ t) S tdt + σ 1 S tdzt 1 dδ t = κ (α δ t) dt + σ 2 dzt 2 where dzt 1dZ t 2 = ρdt. Three-factor model w/ stochastic convenience yield and interest rates: ds t = (r t δ t) S tdt + σ 1 dzt 1 dδ t = κ (ˆα δ t) dt + σ 2 dzt 2 dr t = α (m r) dt + σ 3 dzt 3 where dzt 1dZ t 2 = ρ 1 dt, dzt 2dZ t 3 = ρ 2 dt and dzt 1dZ t 3 = ρ 3 dt.

13 Literature Review (2/2)

14 Literature Review (2/2) Schwartz & Smith 2000 proposes a two factor model w/o a stochastic convenience yield term: S t = exp (χ t + γ t) dχ t = κχ tdt + σ χdz χ t dγ t = µ γdt + σ γdz γ t where dz χ t dz γ t = ρ χγdt. This model is shown to be equivalent to the two-factor model proposed in Schwartz 1997 under linear transformations of the parameters.

15 Literature Review (2/2) Schwartz & Smith 2000 proposes a two factor model w/o a stochastic convenience yield term: S t = exp (χ t + γ t) dχ t = κχ tdt + σ χdz χ t dγ t = µ γdt + σ γdz γ t where dz χ t dz γ t = ρ χγdt. This model is shown to be equivalent to the two-factor model proposed in Schwartz 1997 under linear transformations of the parameters. Hilliard & Reis 1998 extends Schwartz 1997 three-factor model to include jump-diffusion in the commodity spot price in addition to stochastic convenience yield and interest rates. ds t = (µ δ t) dt + σ S dwt S + κdq t S t ) where log (1 + κ) N (log (1 + E [κ]) ω2 2, ω2 and (q t) t 0 is a Poisson counter with intensity λ.

16 Contribution to Literature This work makes two major contributions to literature:

17 Contribution to Literature This work makes two major contributions to literature: 1. Help establish a connection between reduced-form and structural models by endogenously deriving generalized forms of the Schwartz 1997 one-factor and Schwartz & Smith 2000 two-factor models from a simple stochastic dynamic Cournot resource extraction model.

18 Contribution to Literature This work makes two major contributions to literature: 1. Help establish a connection between reduced-form and structural models by endogenously deriving generalized forms of the Schwartz 1997 one-factor and Schwartz & Smith 2000 two-factor models from a simple stochastic dynamic Cournot resource extraction model. 2. Generalize the Cournot model to an arbitrary number of players, N, allowing to derive monopoly, oligopoly and competitive market modes.

19 Model Assumptions Key assumptions underlying the simple model include:

20 Model Assumptions Key assumptions underlying the simple model include: Single common-property resource stock

21 Model Assumptions Key assumptions underlying the simple model include: Single common-property resource stock Costless extraction of resource

22 Model Assumptions Key assumptions underlying the simple model include: Single common-property resource stock Costless extraction of resource No storage

23 Model Assumptions Key assumptions underlying the simple model include: Single common-property resource stock Costless extraction of resource No storage Intertemporally additive discounted utility

24 Model Assumptions Key assumptions underlying the simple model include: Single common-property resource stock Costless extraction of resource No storage Intertemporally additive discounted utility Discounted log() utility maximizing behavior by players

25 Model Assumptions Key assumptions underlying the simple model include: Single common-property resource stock Costless extraction of resource No storage Intertemporally additive discounted utility Discounted log() utility maximizing behavior by players Existence of an inverse demand function (also called the market price function) that maps total quantity extracted by all players to a price

26 Model Assumptions Key assumptions underlying the simple model include: Single common-property resource stock Costless extraction of resource No storage Intertemporally additive discounted utility Discounted log() utility maximizing behavior by players Existence of an inverse demand function (also called the market price function) that maps total quantity extracted by all players to a price Homogenous resource and hence no product differentiation

27 Model Definitions (1/2) Model definitions:

28 Model Definitions (1/2) Model definitions: There are N players

29 Model Definitions (1/2) Model definitions: There are N players Let X t be the resource stock and ɛ t be the demand shock at time t

30 Model Definitions (1/2) Model definitions: There are N players Let X t be the resource stock and ɛ t be the demand shock at time t Let qt i be the resource extracted by player i at time t, then a strategy U i for player i defined as U i : (X t, ɛ t) qt i where qi t Xt t [0, ) is a Markov strategy as it only depends on the current value of the state vector and let U i be the set of such Markov strategies for player i

31 Model Definitions (1/2) Model definitions: There are N players Let X t be the resource stock and ɛ t be the demand shock at time t Let qt i be the resource extracted by player i at time t, then a strategy U i for player i defined as U i : (X t, ɛ t) qt i where qi t Xt t [0, ) is a Markov strategy as it only depends on the current value of the state vector and let U i be the set of such Markov strategies for player i Resource stock is not perfectly measurable, i.e. there is continuous uncertainty regarding the actual stock level, and there are randomly occurring randomly sized jumps in the resource supply, characterizing the evolution of the resource stock as: dx t = ( ) N U i (X t, ɛ t) i=1 ( ) dt + σ X X tdw t + e θt 1 X tdn t where (N t) t 0 is a Poisson process with rate γ that is independent of (W t) t 0. Letting T 1, T 2,... be the arrival times of the Poisson process, the sequence of θ T1, θ T2,... is i.i.d. and θ Ti N (µ θ, σ θ ) i

32 Model Definitions (2/2)

33 Model Definitions (2/2) As long term demand for the resource is linked to stable consumption levels, demand shocks are temporary in nature and should be mean-reverting, characterizing the demand shock process (ɛ t) t 0 as a mean-zero OU process: where dw tdz t = ρdt dɛ t = αɛ tdt + σ ɛdz t

34 Model Definitions (2/2) As long term demand for the resource is linked to stable consumption levels, demand shocks are temporary in nature and should be mean-reverting, characterizing the demand shock process (ɛ t) t 0 as a mean-zero OU process: where dw tdz t = ρdt dɛ t = αɛ tdt + σ ɛdz t The market price function is given by p ( qt 1,..., ) ( N qn t = i=1 qi t ) 1 with p/ q i t < 0

35 Model Definitions (2/2) As long term demand for the resource is linked to stable consumption levels, demand shocks are temporary in nature and should be mean-reverting, characterizing the demand shock process (ɛ t) t 0 as a mean-zero OU process: where dw tdz t = ρdt dɛ t = αɛ tdt + σ ɛdz t The market price function is given by p ( qt 1,..., ) ( N qn t = i=1 qi t ) 1 with p/ q i t < 0 The profit function is given by π ( q i t, p) = pq i t with π/ qi t > 0

36 Problem Formulation Player i s value function Jt i at time t is defined as: [ ( ) ] Jt i (x, ɛ) = E e rs U i (X s, ɛ s) exp (ɛ s) log N t j=1 U ds X t = x, ɛ t = ɛ j (X s, ɛ s) s.t. dx t = ( ) N ( ) U i (X t, ɛ t) dt + σ X X tdw t + e θt 1 X tdn t i=1 dɛ t = αɛ tdt + σ ɛdz t, dw tdz t = ρdt Since (W t) t 0 and (Z t) t 0 are correlated, we can define a new Brownian motion ( ) W t that is independent of both (Wt) t 0 and (Zt) t 0 and represent t 0 dz t = ρdw t + ( 1 ρ 2) 1/2 d Wt. The objective of player i is to find optimal strategy Ui such that: Jt i ( x, ɛ U i, U i ) ( J i t x, ɛ Ui, U i ) Ui U i and t where U i = (U 1,..., U i 1, U i+1,..., U N ). ( ) We define Vt i t x, ɛ Ui, U i as the optimal value function for player i.

37 Solution Overview We take the stochastic dynamic programming approach and write the HJB equation for player i s optimal value function Vt i (x, ɛ) at time t: [ ( )] sup G Vt i (x, ɛ) rv i U i (x, ɛ) exp (ɛ) t (x, ɛ) + log U i U i U i (x, ɛ) + = 0 N j=1,j i U j (x, ɛ) where G is the infinitesimal generator. For brevity, drop the function parameters x, ɛ, then G Vt i is given by PIDE: G Vt i = N U i Uj V t i V i t αɛ x ɛ Vt i 2 x 2 σ2 X x2 j=1,j i Vt i 2 ɛ 2 σ2 ɛ + 2 Vt i [ ] x ɛ σ X σ ɛρx + γe Vt i + V i t where V i t + = Vt ( xe θ, ɛ θ t = θ ) accounting for the jump. We proceed with the solution by fixing and differentiating the HJB equation w.r.t. U i and then looking for a symmetric solution of the type U i = U j i, j [1,..., N].

38 Symmetric Nash Equilibrium Solution

39 Symmetric Nash Equilibrium Solution The optimal value function of player i is: Vt i (x, ɛ) = 1 ( ) 1 r 2 N σ2 X 4r 2 + γ µ θ 2r ( 2r log x r ) N 2 (2N 1) + ɛ α + r

40 Symmetric Nash Equilibrium Solution The optimal value function of player i is: Vt i (x, ɛ) = 1 ( ) 1 r 2 N σ2 X 4r 2 + γ µ θ 2r ( 2r log x r ) N 2 (2N 1) + ɛ α + r The Markov optimal extraction strategy is: U i (x, ɛ) = rx 2N 1 N which is independent of the demand uncertainty ɛ.

41 Symmetric Nash Equilibrium Solution The optimal value function of player i is: Vt i (x, ɛ) = 1 ( ) 1 r 2 N σ2 X 4r 2 + γ µ θ 2r ( 2r log x r ) N 2 (2N 1) + ɛ α + r The Markov optimal extraction strategy is: U i (x, ɛ) = rx 2N 1 N which is independent of the demand uncertainty ɛ. The law of motion of the resource stock under optimal extraction behavior by players is: ( ) dx t = r (2N 1) X tdt + σ X X tdw t + e θt 1 X tdn t

42 Symmetric Nash Equilibrium Solution The optimal value function of player i is: Vt i (x, ɛ) = 1 ( ) 1 r 2 N σ2 X 4r 2 + γ µ θ 2r ( 2r log x r ) N 2 (2N 1) + ɛ α + r The Markov optimal extraction strategy is: U i (x, ɛ) = rx 2N 1 N which is independent of the demand uncertainty ɛ. The law of motion of the resource stock under optimal extraction behavior by players is: ( ) dx t = r (2N 1) X tdt + σ X X tdw t + e θt 1 X tdn t The law of motion of the spot price of the resource is: ( αɛ t r (2N 1) σ2 ɛ 1 2 σɛσ X ρ σ2 X dp t p t = ) dt ( σ ɛρ 1 ) 2 σ ( X dw t + σ ɛ 1 ρ 2 ) ( ) 1/2 d Wt + e 2 1 θt 1 dn t ( ) where Wt and (Wt) t 0 are independent Brownian motions. t 0

43 Key Results (1/2) Derivation of generalized Schwartz 1997 one-factor model If you assume in the Dynamic Cournot model that there is no supply-side uncertainty and the only uncertainty is that of demand uncertainty, then we pick the parameters σ X = 0, θ t = 0 t and ρ = 0. Then, based on the solutions on the prior page, X t = x 0 exp ( rt (2N 1)), with X 0 = x 0 ( ( p t = (rx 0 (2N 1)) 1/2 exp rt N 1 ) ) + ɛ t 2 ( dp t = αɛ t + 1 p t 2 r (2N 1) + 1 ) 2 σ2 ɛ dt + σ ɛd W t Rewriting ɛ t in terms of p t and then plugging into expression for dp t/p t Dynamic Cournot model : dp t = α (µ (t) log (p t)) p tdt + σ ɛp td W t Schwartz one-factor model : ds t = κ (µ log (S t)) S tdt + σs tdz t where µ (t) = r ( N 1 2 ) ( 1 α + t) + σ2 ɛ 2α 1 2 log (rx 0 (2N 1)).

44 Key Results (2/2) Derivation of generalized Schwartz & Smith 2000 two-factor model Taking a detour and looking at the Schwartz & Smith 2000 model, S t = exp (χ t + γ t) dχ t = κχ tdt + σ χdz χ t dγ t = µ γdt + σ γdz γ t, dz χ t dz γ t = ρdt Comparing law of motion of S t to p t from the dynamic Cournot model, dp t p t = ds t S t = dχ t + dγ t (dχt)2 + dχ tdγ t (dγt)2 [ αɛ tdt + σ ɛ (ρdw t + ( 1 ρ 2) 1/2 d W t )] }{{} [( µ X σ2 X dχ t = κχ t dt+σ χdz χ t ) dt 12 σ X dw t ] }{{} dγ t =µ γ dt+σ γ dz γ t [( ) ] + e 1 2 θt 1 dn t } {{ } additional jump term [ σ2 X dt }{{} 1 (dγt )2 2 [ ] σ2 ɛ dt }{{} 1 (dχt )2 2 ] + [ 1 2 σ X σ ɛρdt ] } {{ } dχ t dγ t

45 Limitations of Model and Further Research

46 Limitations of Model and Further Research Limitations

47 Limitations of Model and Further Research Limitations The most obvious limitation of the model is that it considers a single common-property resource stock that is accessible by all players

48 Limitations of Model and Further Research Limitations The most obvious limitation of the model is that it considers a single common-property resource stock that is accessible by all players The model does not consider any extraction costs

49 Limitations of Model and Further Research Limitations The most obvious limitation of the model is that it considers a single common-property resource stock that is accessible by all players The model does not consider any extraction costs The model does not consider any storage

50 Limitations of Model and Further Research Limitations The most obvious limitation of the model is that it considers a single common-property resource stock that is accessible by all players The model does not consider any extraction costs The model does not consider any storage The model does not explicitly consider storage

51 Limitations of Model and Further Research Limitations The most obvious limitation of the model is that it considers a single common-property resource stock that is accessible by all players The model does not consider any extraction costs The model does not consider any storage The model does not explicitly consider storage The log() utility allows for easy decoupling of optimization terms and analytical tractability

52 Limitations of Model and Further Research Limitations The most obvious limitation of the model is that it considers a single common-property resource stock that is accessible by all players The model does not consider any extraction costs The model does not consider any storage The model does not explicitly consider storage The log() utility allows for easy decoupling of optimization terms and analytical tractability Further Research

53 Limitations of Model and Further Research Limitations The most obvious limitation of the model is that it considers a single common-property resource stock that is accessible by all players The model does not consider any extraction costs The model does not consider any storage The model does not explicitly consider storage The log() utility allows for easy decoupling of optimization terms and analytical tractability Further Research Allowing for semi-private resource stocks

54 Limitations of Model and Further Research Limitations The most obvious limitation of the model is that it considers a single common-property resource stock that is accessible by all players The model does not consider any extraction costs The model does not consider any storage The model does not explicitly consider storage The log() utility allows for easy decoupling of optimization terms and analytical tractability Further Research Allowing for semi-private resource stocks Extending the model for geographically dispersed resource stocks and increasing cost of extraction for farther sources

55 Limitations of Model and Further Research Limitations The most obvious limitation of the model is that it considers a single common-property resource stock that is accessible by all players The model does not consider any extraction costs The model does not consider any storage The model does not explicitly consider storage The log() utility allows for easy decoupling of optimization terms and analytical tractability Further Research Allowing for semi-private resource stocks Extending the model for geographically dispersed resource stocks and increasing cost of extraction for farther sources R&D effects and technological improvement

56 Limitations of Model and Further Research Limitations The most obvious limitation of the model is that it considers a single common-property resource stock that is accessible by all players The model does not consider any extraction costs The model does not consider any storage The model does not explicitly consider storage The log() utility allows for easy decoupling of optimization terms and analytical tractability Further Research Allowing for semi-private resource stocks Extending the model for geographically dispersed resource stocks and increasing cost of extraction for farther sources R&D effects and technological improvement Fitting and empirical analysis of the structural model