Multi-Factor Levy Processes and Regime Switching for Commodities: University of Calgary Lunch at the Lab Sebastian Jaimungal sebastian.jaimungal@utoronto.ca Department of Statistics and Mathematical Finance Program, University of Toronto, Toronto, Canada http://www.utstat.utoronto.ca/sjaimung May 19, 2009 (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca1 / 56 Introduction (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca2 / 56 Introduction I Commodities are NOT like equities, interest rates, or currencies! I Commodities are real assets: produced, consumed, transported and stored Customers of commodity derivatives are typically I I I I Industrial producers / consumers Governments Customers tend to be very risk averse due to I I Required consumption Legal risks (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca3 / 56 Crude Oil Forward Price Data Forward price curves from Oct-2006 to Oct-2008 (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca4 / 56 Crude Oil Forward Price Data Daily deviations in Forward price curves Oct-2006, 2007, and 2008 Oct−2006 Forward Price 70 65 60 55 0 1 2 3 4 5 6 4 5 6 4 5 6 Term (years) Oct−2007 Forward Price 85 80 75 70 0 1 2 3 Term (years) Oct−2008 Forward Price 105 100 95 90 85 0 1 2 3 Term (years) (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca5 / 56 Crude Oil Forward Price Data Long end and short end of the Forward price curves from Oct-1998 to Oct-2008 150 Forward Price Short End Middle Long End 100 50 0 3.6 3.65 3.7 3.75 3.8 3.85 3.9 3.95 Time 4 4 x 10 Notice periods of Backwardation and Contango. (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca6 / 56 Crude Oil Forward Price Data Term Structure of volatilities 0.34 0.32 Volatility 0.3 0.28 0.26 0.24 0.22 0.2 0.18 0 1 2 3 4 5 6 7 Term Exponentially decaying volatilities (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca7 / 56 Crude Oil Forward Price Data Correlation Correlation of constant maturity forward price returns 0.95 0.9 0.85 0.8 0.75 0.7 0.65 8 6 8 6 4 4 2 Term 2 0 0 Term As expected, terms further apart less correlated (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca8 / 56 Electricity Data I Electricity data can have huge spikes – but mean revert quickly Figure: Nord Pool Data (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca9 / 56 Electricity Data I Electricity data can have huge spikes – but mean revert quickly Figure: PJM Data (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 10 / 56 Electricity Data I Electricity data can have huge spikes – but mean revert quickly Figure: EW Data (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 11 / 56 Diffusion Models (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 12 / 56 Schwartz’s One-Factor Spot Model I Schwartz(1997) introduced a mean-reverting spot model: dSt = κ(θ − ln St ) St dt + σ St dWt Wt is a Wiener process under the real-world measure P. Wt is a Wiener process under the real-world measure P. I κ controls speed of mean-reversion I θ controls target level I σ controls volatility of paths (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 13 / 56 Schwartz’s One-Factor Spot Model $65 Spott Price $60 $55 k=1 k = 1 $50 k = 10 k = 100 $45 $40 0 0.5 1 1.5 2 2.5 3 Time (years) (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 14 / 56 Schwartz’s One-Factor Spot Model I Choosing market price of risk as λs = λ0 + λ1 ln St maintains mean-reverting model class with new parameters dSt = κ(θ − ln St ) St dt + σ St dW t I Forward prices are then easily obtained Ft (T ) , EQ t [ST ] 0 0 = exp θ + (ln(St ) − θ )e (c) S.Jaimungal (2009) −κ (T −t) σ2 + 1 − e −2κ(T −t) 4κ sebastian.jaimungal@utoronto.ca 15 / 56 Schwartz’s One-Factor Spot Model Sample path of forward price curves: 65 Forward Price 60 55 50 45 40 35 5 4 6 3 4 2 7 5 3 1 Time (years) (c) S.Jaimungal (2009) 0 1 0 2 Term (years) sebastian.jaimungal@utoronto.ca 16 / 56 Two-Factor Spot Models: Pilipovic I Pilipovic(1997) introduced the following model to correct for this (1) dSt = β(θt − St ) dt + σS St dWt (2) dθt = αθt dt + σθ θt dWt (1) with Wt I (2) and Wt uncorrelated Wiener processes I Long-run mean θt blows up I No invariant distribution of commodity spot prices I Does not lead to closed form option prices Xu(2004) generalized this model by incorporating seasonality, making σ time dependent and θt an OU process (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 17 / 56 Two-Factor Spot Models: BGL & HJ I Barlow, Gusev, and Lai(2004) and Hikspoors & J.(2007) introduced a more tractable generalization as follows St = exp{gt + Xt } (1) dXt = α (Yt − Xt ) dt + σ dWt (2) dYt = β (φ − Yt ) dt + η dWt (1) (2) Wt and Wt are correlated Wiener processes, gt incorporates seasonality, and σ can easily be made deterministic (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 18 / 56 Two-Factor Spot Models: BGL & HJ I Sample paths in the two factor model: $100 Spot Price $90 $80 alpha = 0.25 alpha 0.25 $70 $70 alpha = 0.5 $60 alpha = 1 Long Run $50 $40 0 1 2 3 4 5 Term (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 19 / 56 Two-Factor Spot Models: BGL & HJ Sample path of forward price curves in the two-factor BGL & HJ model 70 Forward Price 65 60 55 50 45 5 4 7 6 3 5 4 2 3 1 Time (years) (c) S.Jaimungal (2009) 2 0 1 0 Term (years) sebastian.jaimungal@utoronto.ca 20 / 56 Jump Models (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 21 / 56 One-Factor Jump-Diffusion Model I Clewlow & Strickland (2001) and Cartea & Figueroa (2005) simple extension of Gaussian OU process: St = exp{Xt } dXt = κ(θ − Xt− ) dt + σ dWt + dJt where Jt is a compound Poisson process Jt = Nt X ji n=1 with Nt a Poisson process with activity rate λ and {j1 , j2 , . . . } i.i.d. random variables. (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 22 / 56 One-Factor Jump-Diffusion Model 120 κ = 5 ; σ = 70% κ = 20 ; σ = 95% κ = 100 ; σ = 300% 100 Spot Price 80 60 40 20 0 0 50 100 150 200 250 300 350 400 450 500 Time Notice that as reversion rate increases, need to increase volatility to compensate – otherwise diffusion will be washed out (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 23 / 56 Two-factor Model I Hikspoors & J. (2007) propose instead St = exp{Xt + Yt } dXt = κ(θ − Xt ) dt + σ dWt dYt = −α Yt− dt + dJt where Jt is a pure jump process – such as compound Poisson Jt = Nt X ji n=1 with Nt a Poisson process with activity rate λ and {j1 , j2 , . . . } i.i.d. random variables. (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 24 / 56 Two-factor Model 160 κ=1 κ = 10 κ = 100 140 Spot Price 120 100 80 60 40 20 0 0 50 100 150 200 250 300 350 400 450 500 Time Notice that diffusion and jumps are uncoupled (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 25 / 56 Two-factor Model I Can incorporate frequent small jumps via Lévy processes St = exp{Xt + Yt } dXt = κ(θ − Xt− ) dt + σ dWt + dQt dYt = −α Yt− dt + dJt I Qt – small, frequent, slowly mean-reverting jumps I Carr, Madan, Chang (1998): Variance Gamma (VG) model ν(dy ) = I Carr, Geman, Madan, Yor (2002): CGMY model ν(dy ) = I 1 α y −β|y | e dy ν|y | C −G |y | −M y dy e 1 + e 1 y <0 y >0 |y |1+Y Jt – large, infrequent, fast mean-reverting jumps (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 26 / 56 Multi-Factor Cointegration Model (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 27 / 56 A Multi-Factor Cointegration Model J. & Surkov (2008) introduce the affine multi-factor framework for commodities [à la Duffie, Pan and Singleton (2000) ] dYt = −κYt− dt + dJt , Xt = θ + BYt , St (i) = exp{Xt (i) } I n log spot-prices Xt are modeled as a linear transformation of a set of d−fundamental market factors Yt I θ a d-dimensional vector of long-run means I κ a d × d matrix with positive eigenvalues representing the mixing of the market factors I B a d × n matrix representing the linear transformation of the market factors into the observed log-prices I Jt a d-dimensional Lévy process with Lévy triple (γ, C, ν) (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 28 / 56 A Multi-Factor Model JS Model contains several known models: I One factor mean-reverting model with jumps (Clewlow and Strickland 2000): θ = θ, κ = κ, C = σ 2 , B = 1, and ν(dZ) = λF (dz) dXt = κ(θ − Xt ) dt + σ dWt + dJt (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 29 / 56 A Multi-Factor Model JS Model contains several known models: I Mean-reverting jump-diffusion model (Hikspoors and Jaimungal 2007) with different decay rates for the jumps and diffusion. 2 θ α 0 σ 0 θ= κ= C= B= 1 1 0 0 β 0 0 and ν(dZ1 × dZ2 ) = λ δZ1 dF (Z2 ). dYt1 = α(θ − Yt1 ) dt + σ dWt dYt2 = −βYt2 dt + dJt Xt = Yt1 + Yt2 (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 30 / 56 A Multi-Factor Model JS Model contains several known models: I Two factor mean-reverting model (Barlow, Gusev, and Lai 2004) with log-prices mean-revert to a stochastic long-run mean, which itself mean-reverts to a fixed level: 2 θ α −α σ ρση θ= κ= C= B = 1 0 ν(dZ) = 0 2 θ 0 β ρση η dXt = α(Yt − Xt )dt + σ dWtX dYt = β(θ − Yt )dt + η dWtY (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 31 / 56 A Multi-Factor Model JS model also contains many new models: I Jump-diffusion model with correlated diffusions, idiosyncratic and codependent jumps (via a copula). 2 θ α γ σ ρση 1 0 θ= κ= C= B= φ δ β ρση η 2 0 1 and ν(dZ1 × dZ2 ) = dC (F1 (Z1 ), F2 (Z2 )) with C (u, v ) a copula and Fi (z) two marginal distribution functions. (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 32 / 56 A Multi-Factor Model Allows for independent & codependent jumps... 80 Commodity 1 Commodity 2 75 70 65 Price 60 55 50 45 40 35 30 0 1 2 3 4 5 Time (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 33 / 56 A Multi-Factor Model Allows for cointegration... 90 Raw Refined 1 Refined 2 80 70 Price 60 50 40 30 20 10 0 0.5 1 (c) S.Jaimungal (2009) 1.5 Time 2 2.5 3 sebastian.jaimungal@utoronto.ca 34 / 56 A Multi-Factor Model Allows for cointegration... with jumps too... 70 Raw Refined 1 Refined 2 60 Price 50 40 30 20 10 0 0.5 1 (c) S.Jaimungal (2009) 1.5 Time 2 2.5 3 sebastian.jaimungal@utoronto.ca 35 / 56 A Multi-Factor Model: Induced Forward Prices I Forward prices are martingales under the risk-neutral measure ( (∂t + L)F (t, Y) = 0 , F (T , Y) = exp(θ + BY) , with Lf = (κ(θ − Y))0 ∂Y f + 21 ∂Y0 C∂Y f Z + f (Y + Z) − f (Y) − 1{|Z|<1} Z0 ∂Y f ν(dZ) . Rd /{0} (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 36 / 56 A Multi-Factor Model: Induced Forward Prices I Using the Affine ansatz Fit = exp αit (T ) + β it (T )Yt one finds 0 αit (T ) = θi + Ψ(−i e −κ (T −t) bi , T − t) , 0 β it (T ) = e −κ (T −t) bi , where, (bi )j = Bij . (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 37 / 56 Mean-Reverting Fourier Space Time-Stepping (mrFST) (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 38 / 56 Fourier Space Time-Stepping I Option payoff is given by ϕ(S) I Recall, mean-reverting jump-diffusion process for spot price St = e Xt , I dXt = κ(θ−Xt )dt +σdWt +dJt , X0 = ln S0 Discounted option value v satisfies the PIDE ∂t v + Lv = 0 v (T , x) = ϕ(e x ) where L is the infinitesimal generator: 1 Lf (x) = κ(θ − x)∂x f (x) + σ 2 ∂xx f (x) 2 Z + (f (x + y ) − f (x)) ν(dy ) (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 39 / 56 Fourier Space Time-Stepping I The conditional characteristic function under Q is ht,x (ω, T ) , EQ [e iωXT |Ft ] I Assume the affine form: T (ω)+ΦT (ω)x t ht,x (ω, T ) = e Ψt I T ΨT t (ω) and Φt (ω) then satisfy a system of Riccati ODEs: ( R ΦT y 1 2 T 2 T t − 1)ν(dy ) = 0 Ψ̇T t + κθΦt + 2 σ (Φt ) + (e T T Φ̇t − κΦt = 0, T subject to ΨT T (ω) = 0 and ΦT (ω) = iω (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 40 / 56 Fourier Space Time-Stepping I Can solve for ΦT t (ω) analytically: −κ(T −t) ΦT t (ω) = iωe I ΨT t (ω) can then be solved: ω2σ2 −κ(T−t) (1−e −2κ(T−t) ) ΨT (ω) = iωθ(1−e ) − t 4κ Z T + ψ̃(ωe −k(T−s) )ds t where ψ̃ is the characteristic function of the jump distribution I Last term can be computed analytically for double-exponential distribution (Kou model) and numerically using quadratures for others (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 41 / 56 Fourier Space Time-Stepping I Expand the payoff (assume paid at t + ∆t) in a Fourier basis: Z ∞ 1 e iωx F[vt+∆t ](ω) dω ϕ(x) = vt+∆t (x) = 2π −∞ I Assuming no decisions (such as barrier breach or optimal exercise) are made during the interval (t, t + ∆t]: Z ∞ 1 vt (x) = ht,x (ω, t + ∆t) F[vt+∆t ](ω) dω 2π −∞ The above satisfies the PIDE and the boundary condition I Apply Fourier transform to vt (x): Z ∞Z ∞ t+∆t t+∆t dx 0 0 F[vt ](ω) = e Ψt (ω )+Φt (ω )x F[vt+∆t ](ω 0 )dω 0 e −iωx 2π −∞ −∞ (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 42 / 56 Fourier Space Time-Stepping The price Vt (Yt ) of a European option written on the vector of (n) (1) (1) price processes {St = e Xt , . . . , Stn = e Xt } with payoff function ϕ(ST ) = ϕ(θ e + B YT ) = φ(YT ) is h i 0 Vt (Yt ) = F −1 F[φ(YT )](e κ (T −t) ω) e Ψ(ω,T −t)+(T −t)Tr κ (Yt ) , where, Z s 0 ψ(e κ u ω) du , and 0 Z 0 e iω y − 1 − i 1{|y|<1} ω 0 y ν(dy) , ψ(ω) = − 12 ω 0 Cω + Ψ(ω, s) = Rd /{0} F[.] and F −1 [.] denote the Fourier and inverse Fourier transforms respectively, and 0 denotes transpose. (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 43 / 56 Fourier Space Time-Stepping I Mean-Reverting FST Method h i κ0 ∆t ) 0 v m−1 (Y) = FFT−1 e Ψ(ωe · FFT[v m (Y0 e −κ ∆t )] I Without mean reversion, mrFST reduces to the standard FST method of Jackson, Jaimungal and Surkov (2008) I For Bermudan style claims: h i o n em ](ω) e Ψ(ω,∆tm ) (Y) , VM (Y) . Vm−1 (Y) = max e −r ∆tm F−1 F[V A penalty method can also be used. I For Barrier style claims: h i em ](ω) e Ψ(ω,∆tm ) (Y) · I{θ+BY<B} Vm−1 (Y) =e −r ∆tm F−1 F[V + R · I{θ+BY≥B} . (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 44 / 56 Numerical Examples (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 45 / 56 Fourier Space Time-Stepping Comparison of Tree, LSMC, and mrFST for mean-reverting diffusion: American Put Option 50 LSMC mrFST Tree 49.5 49 Asset Price 48.5 48 47.5 47 46.5 46 45.5 45 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Time (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 46 / 56 Fourier Space Time-Stepping Comparison of LSMC and mrFST for mean-reverting diffusion with Kou jumps: American Put Option 50 LSMC mrFST 49 Asset Price 48 47 46 45 44 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Time (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 47 / 56 Fourier Space Time-Stepping Sensitivity of boundary on asymmetry using mrFST mean-reverting diffusion with Kou jumps: American Put Option 50 p=0.5 p=0.4 p=0.3 p=0.2 49 Asset Price 48 47 46 45 44 43 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Time (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 48 / 56 Swing Options I Common in electricity and natural gas markets I Provides constrained flexibility with respect to volume and timing of energy delivery I Two components: a pure forward agreement and a swing option An Example I I I I I I The holder agrees to buy 100MWh at a cost of $45/MWh over a period of 1 month. At the start of each hour, the holder has the right to increase power consumption to 110MW for that hour (swing up) or decrease to 90MW (swing down) at the same cost. The total number of swings is limited to 50. The swing component is the right to change consumption at holder’s choosing. For overview of swing option valuation see Ware (2007) (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 49 / 56 Swing Options At each swing opportunity, a choice to exercise q swing “options” for immediate cashflow Υ must be made: Dynamic Programming Equation n o vtm (x, Q) = max Υtm (x, q) + e −r ∆t E[vtm+1 (x, Q + q)] q where the expectation is readily computed using the mrFST method I I I The available choices are to do nothing q = 0, swing up q > 0 or swing down q < 0 The amount of swings mayPbe bounded |Qtm | ≤ Q where P Qtm = m q |qtj | j=1 tj or Qtm = The cashflow function Υt (x, q) may include a penalty term to enforce additional limits on Q or may be as simple as Υt (x, q) = q(e x − K ) (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 50 / 56 Swing Options 230.0 Swing Value (V) 210.0 =0.05 =0.25 =0.5 =1.25 190.0 170.0 150.0 130.0 110.0 40.0 55.0 70.0 85.0 100.0 115.0 130.0 Spot Price (S) I I Option: Swing S = 100, K = 100, T = 2, −3 ≤ Q ≤ 5 Model: Kou jump-diffusion with mean reversion σ = 75%, λ = 0.5, p = 0.45, η+ = 0.25, η− = 0.2, θ = 80, r = 0.05 (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 51 / 56 Real Option – Regime Switching Volume I I Consider the option to invest with limited lifetime Volume is not known, but you have estimates which are continually updated according to a Markov chain Zt a exp(‐b t) V1 I V2 V3 V4 V5 Project Value Pt is then modeled as Pt = VZt exp{Xt } dXt = κ(θ − Xt ) dt + σdWt I Investment amount is I . Option to invest is Bermudan style: ft = sup E e −r τ VZτ e Xτ − I τ ∈T (c) S.Jaimungal (2009) Ft + sebastian.jaimungal@utoronto.ca 52 / 56 Real Option – Regime Switching Volume I Between exercise dates, can show that the price f (i) (t, x) contingent on Zt = i satisfies system of coupled PDEs: ( P (j) = 0 (∂t + L − r )f (i) + ht m j=1 Aij f f (i) (tn−1 , x) = max(f (i) (tn+ , x) , ϕ(i) (x)) where ϕ(i) (x) = (Vi e x − I )+ I An mrFST regime switching algorithm can be worked out: i h n f · (tn−1 , x) = max e −r ∆tn F−1 Hn F[e fn ](ω) e Ψ(ω,∆tm ) (x) , (Vi e x − I )+ where Hn = Ue Λ (c) S.Jaimungal (2009) R tn tn −1 hs ds U −1 sebastian.jaimungal@utoronto.ca 53 / 56 Real Option – Regime Switching Volume Comparison of constant versus Markov switching Trigger curves (Value / Investment) vol=0.6 vol=0.8 vol=1 vol=1.2 vol=1.4 2.6 Value / Investment 2.4 2.2 2 1.8 1.6 1.4 1.2 1 0 2 4 6 8 10 Time (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 54 / 56 Real Option – Regime Switching Volume Comparison of constant versus Markov switching Trigger curves (Value / Investment) – SLOWER “learning” vol=0.6 vol=0.8 vol=1 vol=1.2 vol=1.4 2.6 Value / Investment 2.4 2.2 2 1.8 1.6 1.4 1.2 1 0 2 4 6 8 10 Time (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 55 / 56 THE END!! THANK YOU FOR YOUR ATTENTION!! ANY QUESTIONS? sebastian.jaimungal@utoronto.ca http://www.utstat.utoronto.ca/sjaimung (c) S.Jaimungal (2009) sebastian.jaimungal@utoronto.ca 56 / 56