Financial modeling with jump processes is a fully explored domain. We present in this post some of the theoretical aspects of Lévy processes.

We also expose a subset of jump diffusion models available into LexiFi Apropos. Underlying dynamics, European call pricing and calibration procedures are developed for the Merton and Bates models.
For the sake of clarity, some theoretical aspects and figures are hidden, don’t hesitate to expand them for having more information.

Element of stochastic calculus

In this section, we recall some key points on the extension to discontinuous processes (Itô formula, measure theory, … ). we also present Lévy processes (Lévy-Itô decomposition, Lévy-Khintchine representation, …). The reader can expand/collapse for more information

Extension to discontinuous processes

Generalised Itô formula

Let (Xt)t∈R+ (X_t)_{t\in\mathbb{R}^+} be a cadlag (right-continuity and left limits) process such that XX is a continuous semi-martingale and ff a twice differentiable function. Continuous Itô formula:

f(Xt)=f(X0)+∫0tf′(Xs)dXs+12∫0tf′′(Xs)d[X]s f(X_{t}) = f(X_{0}) + \int\limits_{0}^{t} {f'(X_{s}) dX_{s} } + \frac{1}{2} \int\limits_{0}^{t} {f''(X_{s}) d \left[ X \right]_{s} }

Generalised Itô formula: Let X X be a cadlag process.

f(Xt)=f(X0)+∫0tf′(Xs−)dXs+12∫0tf′′(Xs−)d[X]s+∑s≤t[△f(Xs)−f′(Xs−)△Xs−−12f′′(Xs−)△Xs2] \begin{align*} f(X_{t}) = f(X_{0}) & + \displaystyle \int\limits_{0}^{t} {f'(X_{s^{-}}) dX_{s} } + \frac{1}{2} \int\limits_{0}^{t} {f''(X_{s^-}) d \left[ X \right]_{s} } \\ & + \displaystyle \sum_{s \le t} {\left[ \triangle f(X_{s}) - f'(X_{s^{-}}) \triangle X_{s^{-}} - \frac{1}{2} f''(X_{s^-}) \triangle X_{s}^{2} \right]} \end{align*}

with

  • Xt−=lim⁡s→ts<tXs X_{t^{-}} = \displaystyle \lim_{s \to t \atop s < t} X_s
  • △Xt=Xt−Xt− \triangle X_t = X_t - X_{t^{-}}
  • ∑s≤t△Xs=∑{ti,0≤ti≤t    and    △Xti≠0}△Xti \displaystyle \sum_{s \le t} \triangle X_s = \displaystyle \sum_{ \{ t_i , 0 \le t_i \le t \;\; and \;\; \triangle X_{t_i} \neq 0 \} } \triangle X_{t_i} such that {ti,0≤ti≤t \{ t_i , 0 \le t_i \le t and △Xti≠0} \triangle X_{t_i} \neq 0 \} is countable

Splitting: continuous/discontinuous processes

Decomposition (continuous and discontinuous):

X=Xc⏟continuous  part+Xd⏟discontinuous  part=Xc+∑s≤.△Xs X = \underbrace{X^{c}}_{continuous \; part} + \underbrace{X^{d}}_{discontinuous \; part} = X^{c} + \displaystyle \sum_{s \le .} \triangle X_{s}

The quadratic covariation between two processes is defined by:

[X,Y]t=lim⁡∣σn∣→0∑i=1n(Xti−Xti−1)(Yti−Yti−1) \left[ X,Y \right]_{t} = \displaystyle \lim_{ \left| \sigma_{n} \right| \to 0} \sum_{i=1}^{n} (X_{t_i} - X_{t_{i-1}}) (Y_{t_i} - Y_{t_{i-1}})

with ∣σn∣=sup⁡{ti−ti−1∀i∈1,…,n} \left| \sigma_{n} \right| = \sup \{ t_i -t_{i-1} \forall i \in 1,\ldots ,n \} and σn={t0=0<…<tn=t} \sigma_n = \{ t_0 = 0 < \ldots < t_n = t \} a subdivision of [0;t] [ 0 ; t ] (size n + 1).

We expose two interesting features:

  • [X,Y]tc=[X,Y]t−∑s≤t△Xs△Ys=[Xc,Yc]t [X,Y]_{t}^{c} = [X,Y]_{t} - \displaystyle \sum_{s \le t} \triangle X_{s} \triangle Y_{s} = [X^{c},Y^{c}]_{t}
  • dXt=dXtc+△Xt d X_t = d X_t^c + \triangle X_t

Lévy process

Definition

Let XX be a stochastic process. XX is a Lévy process if

  • X0=0X_0 = 0 and XX cadlag process
  • XX has independent and stationary increments

Two important features are:

  • XX is continuous in probability: ∀ϵ∈R+,lim⁡s→0P(∣Xs+t−Xt∣>ϵ)=0\forall \epsilon \in \mathbb{R}^{+}, \displaystyle \lim_{s \rightarrow 0} \mathbb{P} ( |X_{s+t} - X_t| > \epsilon ) = 0
  • XX has no jump at a deterministic time: ∀t∈R+,P(Xt−=Xt)=1 \forall t \in \mathbb{R}^+, \mathbb{P} (X_{t^-} = X_t) = 1

Remark: If XX is a continuous Lévy process, it can be written as follow: Xt=γt+AWt X_t = \gamma t + A W_t with γ∈Rd\gamma \in \mathbb{R}^d ,WW a d-dimensional Brownian motion and AA a positive semi-definite matrix.

Probability measure

Let (Ω,P,F) (\Omega, P, \mathcal{F}) be a probability space and (E,ε)(E, \varepsilon) a measurable set.

Random measure:

MM: Ω×ε→R\Omega \times \varepsilon \rightarrow \mathbb{R} is a random measure if

  • ∀ω∈Ω\forall \omega \in \Omega, M(ω,⋅)M(\omega,\cdot) is a measure on ε\varepsilon.
  • ∀A∈ε\forall A \in \varepsilon, M(⋅,A)M(\cdot,A) is measurable.

Poisson random measure:

Let μ\mu be a measure on (E,ε)(E, \varepsilon), MM: Ω×ε→R\Omega \times \varepsilon \rightarrow \mathbb{R} is a Poisson random measure with intensity μ\mu if

  • ∀A∈ε\forall A \in \varepsilon with μ(A)<+∞\mu(A) < +\infty , M(⋅,A)M(\cdot,A) follows a Poisson law with intensity μ(A) \mu(A) ( =E[M(⋅,A)]  )( \, = E[M(\cdot,A)] \,\,).
  • ∀(A1,…,An)∈ϵn\forall (A_1, \ldots, A_n) \in \epsilon^n separated sets, M(⋅,A1),…,M(⋅,An)M(\cdot,A_1), \ldots, M(\cdot,A_n) are independents.

Jump measure:

Let X be a cadlag process with value in Rd\mathbb{R}^d. The jump measure of XX is a random measure on [0;∞)×Rd [0;\infty) \times \mathbb{R}^d defined by:

∀A∈(R+×Rd),JX(A)=#{t:ΔXt≠0 and (t,ΔXt)∈A} \forall A \in ( \mathbb{R}^+ \times \mathbb{R}^d),\quad J_X(A) = \# \{t:\Delta X_t \neq 0 \text{ and } (t,\Delta X_t) \in A \}

Property: A jump measure of a Lévy process is a Poisson measure.

Remark: A jump measure of the set ([t1,t2],B)([t1,t2],B) corresponds to the number of jump between t1t1 and t2t2 with a size include in BB.

Lévy measure:

Let XX be a Lévy process with value in Rd\mathbb{R}^d. The Lévy measure of XX is a measure defined by:

∀A∈B(Rd),ν(A)=E[#{t∈[0,1]:ΔXt≠0 and ΔXt∈A}] \forall A \in \mathcal{B} (\mathbb{R}^d),\quad \nu(A) = \mathbb{E}[ \# \{ t \in [0,1]: \Delta X_t \neq 0 \text{ and }\Delta X_t \in A \} ]

Lévy-Itô decomposition

Let XX be a Lévy process with value in Rd\mathbb{R}^d and the Lévy measure ν\nu then

  • Its jump measure JX J_X is a Poisson random measure on [0;∞)[0;\infty) with intensity dt×νdt \times \nu
  • Its Lévy measure verifies ∫Rd(∥x∥2)∧1)ν(dx)<∞ \displaystyle \int\limits_{\mathbb{R}^d} (\|x\|^2)\wedge 1) \nu(dx) < \infty
  • There exists γ∈Rd\gamma \in \mathbb{R}^d and a d-dimensional brownian motion BB with covariance matrix AA such that: Xt=γt+Bt+Nt+MtNt=∫∣x∣>1,s∈[0,t]xJX(ds×dx)Mt=∫0<∣x∣<1,s∈[0,t]xJ~X(ds×dx)=∫0<∣x∣<1,s∈[0,t]x[JX(ds×dx)−ν(dx)ds] \begin{align*} X_t = & \gamma t + B_t + N_t + M_t \\ N_t = & \displaystyle \int\limits_{ |x| > 1, s \in [0,t] } x J_X(ds \times dx) \\ M_t = & \displaystyle \int\limits_{ 0 < |x| < 1, s \in [0,t] } x \tilde{J}_X(ds \times dx) = \displaystyle \int\limits_{ 0 < |x| < 1, s \in [0,t] } x [J_X(ds \times dx) - \nu(dx)ds] \end{align*}

The process NN corresponds to a composed Poisson process with jump greater than 1.

The process MM corresponds to a composed and compensated Poisson process with jump lower than 1.

We remove the mean of small jumps so that the integral will be computable (can be infinite). Let’s note that MM is a martingale.

We only have to define a Lévy process using the following tuple: (A,ν,γ)(A, \nu, \gamma). This is called the characteristic triplet.

Remark: We find the commonly used formula for XX a continuous Lévy process. The characteristic triplet is (A,0,γ)(A, 0, \gamma ). We can rewrite the previous decomposition with ϵ∈R \epsilon \in \mathbb{R} :

Nt=∫∣x∣>ϵ,s∈[0,t]xJX(ds×dx)Mt=∫0<∣x∣<ϵ,s∈[0,t]xJ~X(ds×dx)=∫0<∣x∣<ϵ,s∈[0,t]x[JX(ds×dx)−ν(dx)ds]γϵ=γ−∫ϵ<∣x∣<1,s∈[0,t]xν(dx)ds \begin{align*} N_t = & \displaystyle \int\limits_{ |x| > \epsilon, s \in [0,t] } x J_X(ds \times dx) \\ M_t = & \displaystyle \int\limits_{ 0 < |x| < \epsilon, s \in [0,t] } x \tilde{J}_X(ds \times dx) = \displaystyle \int\limits_{ 0 < |x| < \epsilon, s \in [0,t] } x [J_X(ds \times dx) - \nu(dx)ds] \\ \gamma^\epsilon = & \gamma - \displaystyle \int\limits_{ \epsilon < |x| < 1, s \in [0,t] } x \nu(dx)ds \end{align*}

We can compute the lim⁡ϵ→0\displaystyle \lim_{\epsilon \rightarrow 0} in the decomposition formula.

Lévy-Khintchine representation

Let XX be a Lévy process with characteristic triplet (A,ν,γ)(A, \nu, \gamma). The characteristic function is:

E[eiuXt]=exp{ t [ i γ u−Au22+∫R(ei u x−1−i u x1∣x∣≤1)ν(dx) ] } \mathbb{E} [ e^{iuX_t} ] = exp \left\{\, t \, \big[\, i\, \gamma\, u - \frac{A u^2}{2} + \displaystyle \int\limits_{\mathbb{R}} ( e^{i \, u \, x} - 1 - i \, u \, x 1_{|x| \le 1} ) \nu (dx) \,\big]\, \right\}

Exponential formula

Let M be a Poisson random measure on (E,ε)(E, \varepsilon) with intensity μ\mu, B∈εB \in \varepsilon and f a measurable function such that ∫B∣ef(x)−1∣μ(dx)<∞\displaystyle \int\limits_{B} |e^{f(x)} - 1| \mu(dx) < \infty. The following equality holds:

E[e ∫Bf(x)M(dx) ]=exp[ ∫B(ef(x)−1)μ(dx) ] \mathbb{E} \left[e^{ \, \int\limits_{B} f(x) M(dx) \,}\right] = exp \left[ \, \int\limits_{B} ( e^{f(x)} - 1 ) \mu(dx)\,\right]

Exponential Lévy process

Let XX be a Lévy process with triplet (σ,ν,γ) (\sigma, \nu, \gamma) . Let’s note Y=eXY = e^{X} . We can apply the generalise Itô formula for semi-martingale:

Yt=eX0+∫0teXs−dXs+12∫0teXs−d[Xs]+∑s≤t[ΔeXs−eXs−ΔXs ]=1+∫0tYs−dXs+12∫0tYs−σ2ds+∑s≤t[eXs−+ΔXs−eXs−−eXs−ΔXs]=1+∫0tYs−dXs+12∫0tYs−σ2ds+∑s≤tYs−[eΔXs−1−ΔXs]=1+∫0tYs−dXs+12∫0tYs−σ2ds+∫0t∫RYs−[ex−1−x]JX(ds×dx)Yt=1+∫0tYs−[γ+σ22]ds+∫0tYs−σdWs+∫0tYs−∫RxJX(ds×dx)−∫0tYs−∫−11xν(dx)ds+∫0t∫R[Ys−[ex−1−x]JX(ds×dx) \begin{align*} Y_t = & e^{X_0} + \int\limits_0^t e^{X_{s^-}} d X_s + \frac{1}{2} \int\limits_0^t e^{X_{s^-}} d [X_s] + \displaystyle \sum\limits_{s \le t} [ \Delta e^{X_s} - e^{X_{s^-}}\Delta {X_s} \, ] \\ = & 1 + \int\limits_0^t Y_{s^-} dX_s + \frac{1}{2} \int\limits_0^t Y_{s^-} \sigma^2 ds + \displaystyle \sum\limits_{s \le t} [ e^{X_{s^-} + \Delta X_s} - e^{X_{s^-}} - e^{X_{s^-}} \Delta X_s ] \\ = & 1 + \int\limits_0^t Y_{s^-} dX_s + \frac{1}{2} \int\limits_0^t Y_{s^-} \sigma^2 ds + \displaystyle \sum\limits_{s \le t} Y_{s^-} [ e^{\Delta X_s} - 1 - \Delta X_s ] \\ = & 1 + \int\limits_0^t Y_{s^-} dX_s + \frac{1}{2} \int\limits_0^t Y_{s^-} \sigma^2 ds + \int\limits_0^t \int\limits_{\mathbb{R}} Y_{s^-} [ e^{x} - 1 - x ] J_X (ds \times dx) \\ Y_t = & 1 + \int\limits_0^t Y_{s^-} [ \gamma + \frac{\sigma^2}{2}] ds + \int\limits_0^t Y_{s^-} \sigma dW_s \\ & + \int\limits_0^t Y_{s^-} \int\limits_{\mathbb{R}} x J_X (ds \times dx) - \int\limits_0^t Y_{s^-} \int\limits_{-1}^{1} x \nu (dx) ds \\ & + \int\limits_0^t \int\limits_{\mathbb{R}} [ Y_{s^-} [ e^{x} - 1 - x ] J_X (ds \times dx) \end{align*}

We can split the process in a martingale part and a finite variation part. We have:

Yt=1+∫0tYs−[γ+σ22]ds+∫0tYs−σdWs−∫0tYs−∫−11xν(dx)ds+∫0t∫RYs−[ex−1]JX(ds×dx)+∫0t∫RYs−[ex−1]ν(dx)ds−∫0t∫RYs−[ex−1]ν(dx)dsYt=1+∫0tYs−[γ+σ22]ds+∫0tYs−σdWs−∫0tYs−∫−11xν(dx)ds+∫0t∫RYs−[ex−1]J~X(ds×dx)+∫0t∫RYs−[ex−1]ν(dx)dsYt=1+∫0tYs−[γ+σ22]ds+∫0tYs−σdWs+∫0t∫RYs−[ex−1]J~X(ds×dx)+∫0t∫RYs−[ex−1−x1∣x∣<1]ν(dx)ds \begin{align*} Y_t = & 1 + \int\limits_0^t Y_{s^-} [ \gamma + \frac{\sigma^2}{2}] ds + \int\limits_0^t Y_{s^-} \sigma dW_s - \int\limits_0^t Y_{s^-} \int\limits_{-1}^{1} x \nu (dx) ds \\ & + \int\limits_0^t \int\limits_{\mathbb{R}} Y_{s^-} [ e^{x} - 1] J_X (ds \times dx) + \int\limits_0^t \int\limits_{\mathbb{R}} Y_{s^-} [ e^{x} - 1] \nu (dx) ds - \int\limits_0^t \int\limits_{\mathbb{R}} Y_{s^-} [ e^{x} - 1] \nu (dx) ds \\ Y_t = & 1 + \int\limits_0^t Y_{s^-} [ \gamma + \frac{\sigma^2}{2}] ds + \int\limits_0^t Y_{s^-} \sigma dW_s - \int\limits_0^t Y_{s^-} \int\limits_{-1}^{1} x \nu (dx) ds \\ & + \int\limits_0^t \int\limits_{\mathbb{R}} Y_{s^-} [ e^{x} - 1 ] \tilde{J}_X (ds \times dx) + \int\limits_0^t \int\limits_{\mathbb{R}} Y_{s^-} [ e^{x} - 1] \nu (dx) ds \\ Y_t = & 1 + \int\limits_0^t Y_{s^-} [ \gamma + \frac{\sigma^2}{2}] ds + \int\limits_0^t Y_{s^-} \sigma dW_s \\ & + \int\limits_0^t \int\limits_{\mathbb{R}} Y_{s^-} [ e^{x} - 1 ] \tilde{J}_X (ds \times dx) + \int\limits_0^t \int\limits_{\mathbb{R}} Y_{s^-} [ e^{x} - 1 - x 1_{|x| < 1}] \nu (dx) ds \end{align*}

We therefore have Y=eX=A+M Y = e^{X} = A + M , with A A and M M defined as follows:

∀t∈R+,At=∫0tYs−[γ+σ22]ds+∫0t∫R[ex−1−x1∣x∣<1]ν(dx)]ds∀t∈R+,Mt=1+∫0tYs−σdWs+∫0t∫RYs−[ex−1]J~X(ds×dx) \begin{align*} \forall t \in \mathbb{R}^+, A_t = & \int\limits_0^t Y_{s^-} [ \gamma + \frac{\sigma^2}{2} ] ds + \int\limits_0^t \int\limits_{\mathbb{R}} [ e^{x} - 1 - x 1_{|x| < 1}] \nu (dx) ] ds \\ \forall t \in \mathbb{R}^+, M_t = & 1 + \int\limits_0^t Y_{s^-} \sigma dW_s + \int\limits_0^t \int\limits_{\mathbb{R}} Y_{s^-} [ e^{x} - 1 ] \tilde{J}_X (ds \times dx) \end{align*}

Generality on exponential Lévy models

We develop in this section a general dynamic for exp-Lévy models. We expose a valuation method using partial integro-differential equation (PIDE). The reader can expand/collapse for more information.

Underlying dynamic

Let StS_t be a random process standing for the underlying price under a risk neutral probability.

∀t∈R,St=ert+Xt \forall t \in \mathbb{R}, S_t = e^{rt + X_t}

with XtX_t a Lévy process with characteristic triplet (σ,ν,γ)(\sigma, \nu, \gamma)

Let’s note ∀t∈R+ \forall t \in \mathbb{R}^+ , C(t)=e∫0tr(s)ds C(t) = e^{ \int\limits_0^t r(s)ds } , ⇔ \Leftrightarrow dC(t)C(t−)=rdt \frac{d C(t)}{C(t^-)} = r dt ⇔ \Leftrightarrow C(t)=∫0trC(s−)ds C(t) = \int\limits_0^t r C(s^{-}) ds

We have

∀t∈R,S~t=StCt=eXt=At+Mt \forall t \in \mathbb{R}, \tilde{S}_t = \frac{S_t}{C_t} = e^{X_t} = A_t + M_t

with At A_t and Mt M_t defined in the previous decomposition formula.

Under the risk neutral probability, the actualized future price must be a martingale. We can write:

dS~t=d(At+Mt)=dAt+dMt d\tilde{S}_t = d( A_t + M_t ) = dA_t + dM_t

Therefore, we must have dAt=0 dA_t = 0 and then:

γ+σ22+∫R[ex−1−x1∣x∣<1]ν(dx)=0 \gamma + \frac{\sigma^2}{2} + \displaystyle \int\limits_{\mathbb{R}} [ e^{x} - 1 - x 1_{|x| < 1}] \nu (dx) = 0

The dynamic of the actualized underlying is:

dS~t=dMt=St− [ σdWt+∫R[ex−1]J~X(dt×dx) ] d\tilde{S}_t = dM_t =S_{t^-} \, [ \, \sigma d W_t + \int\limits_{\mathbb{R}} [ e^{x} - 1 ] \tilde{J}_X (dt \times dx) \, ]

The dynamic is:

dStSt−=rdt+σdWt+∫R(ex−1)J~X(dt×dx) \frac{dS_t}{S_t^-} = r dt + \sigma dW_t + \int\limits_{\mathbb{R}} (e^x - 1) \tilde{J}_X (dt \times dx)

with

  • r,σ∈R2 r, \sigma \in \mathbb{R}^2
  • J~X(ds×dx)=JX(ds×dx)−ν(dx)ds \tilde{J}_X (ds \times dx) = J_X (ds \times dx) - \nu(dx) ds
  • ν \nu the Lévy measure of XX
  • JX J_X the random Poisson measure on R+\mathbb{R}^+ with intensity dt×νdt \times \nu

Partial integro-differential equation

Let’s note ∀t∈R,V(t,St)\forall t \in \mathbb{R}, V(t,S_t) the price of a contingent claim.

To find the partial integro-differential equation, let’s notice that under the risk neutral probability the actualised future price must be a martingale. Applying Itô formula, ∀t∈R\forall t \in \mathbb{R},

d(e−rtV(t,St))=−re−rtV(t,St)dt+e−rt∂V(t,St)∂tdt+e−rt∂V(t,St)∂StdSt+e−rt∂2V(t,St)∂St  2σ22dt+e−rtΔV(t,St)−e−rt∂V(t,St)∂StΔSt \begin{align*} d( e^{-rt} V(t,S_t)) = & -r e^{-rt}V(t,S_t) dt + e^{-rt} \frac{\partial V(t,S_t)}{\partial t}dt + e^{-rt}\frac{\partial V(t,S_t)}{\partial S_t} dS_t \\ & + e^{-rt}\frac{\partial^2 V(t,S_t)}{\partial S_t^{\,\,2}} \frac{\sigma^2}{2} dt + e^{-rt} \Delta V(t,S_t) - e^{-rt} \frac{\partial V(t,S_t)}{\partial S_t}\Delta S_t \end{align*}

With:

dSt=St−[rdt+σdWt+∫R(ex−1)J~X(dt×dx)]ΔSt=St−eΔXt−St−=∫RSt−(ex−1)JX(dt×dx)ΔV(t,St)=V(t,St−eΔXt)−V(t,St−)=∫R[V(t,St−ex)−V(t,St−)]JX(dt×dx) \begin{align*} dS_t & = S_t^- \left[ r dt + \sigma dW_t + \int\limits_{\mathbb{R}} (e^x - 1) \tilde{J}_X (dt \times dx) \right] \\ \Delta S_t & = S_{t^-}e^{\Delta X_t} - S_{t^-} = \int\limits_{\mathbb{R}} S_{t^-} (e^x - 1) J_X (dt \times dx) \\ \Delta V(t,S_t) & = V(t,S_{t^-}e^{\Delta X_t} ) - V(t,S_{t^-}) = \int\limits_{\mathbb{R}} [ V(t,S_{t^-}e^x ) - V(t,S_{t^-}) ] J_X (dt \times dx) \end{align*}

By replacing:

d(e−rtV(t,St))=−re−rtV(t,St)dt+e−rt∂V(t,St)∂tdt+e−rt∂V(t,St)∂StSt−[rdt+σdWt+∫R(ex−1)J~X(dt×dx)]+e−rt∂2V(t,St)∂St  2σ22dt+e−rt∫R[V(t,St−ex)−V(t,St−)]JX(dx×dt)−e−rt∂V(t,St)∂St∫RSt−(ex−1)JX(dt×dx)d(e−rtV(t,St))=−re−rtV(t,St)dt+e−rt∂V(t,St)∂tdt+e−rt∂V(t,St)∂StSt−rdt+e−rt∂V(t,St)∂StSt−σdWt+e−rt∂V(t,St)∂StSt−∫R(ex−1)JX(dt×dx)−e−rt∂V(t,St)∂StSt−∫R(ex−1)ν(dt)dx+e−rt∂2V(t,St)∂St  2σ22dt+e−rt∫R[V(t,St−ex)−V(t,St−)]JX(dt×dx)−e−rt∂V(t,St)∂St∫RSt−(ex−1)JX(dt×dx)d(e−rtV(t,St))=−re−rtV(t,St)dt+e−rt∂V(t,St)∂tdt+e−rt∂V(t,St)∂StSt−rdt+e−rt∂V(t,St)∂StSt−σdWt−e−rt∂V(t,St)∂StSt−∫R(ex−1)ν(dt)dx+e−rt∂2V(t,St)∂St  2σ22dt+e−rt∫R[V(t,St−ex)−V(t,St−)]J~X(dt×dx)+e−rt∫R[V(t,St−ex)−V(t,St−)]ν(dx)dt \begin{align*} d( e^{-rt} V(t,S_t)) = & -r e^{-rt}V(t,S_t) dt + e^{-rt} \frac{\partial V(t,S_t)}{\partial t}dt + e^{-rt}\frac{\partial V(t,S_t)}{\partial S_t} S_t^-[r dt + \sigma dW_t \\ & + \int\limits_{\mathbb{R}} (e^x - 1) \tilde{J}_X (dt \times dx)] + e^{-rt}\frac{\partial^2 V(t,S_t)}{\partial S_t^{\,\,2}} \frac{\sigma^2}{2} dt \\ & + e^{-rt} \int\limits_{\mathbb{R}} [ V(t,S_{t^-}e^x ) - V(t,S_{t^-}) ] J_X (dx \times dt) \\ & - e^{-rt} \frac{\partial V(t,S_t)}{\partial S_t}\int\limits_{\mathbb{R}} S_{t^-} (e^x - 1) J_X (dt \times dx) \\ d( e^{-rt} V(t,S_t))= & -r e^{-rt}V(t,S_t) dt + e^{-rt} \frac{\partial V(t,S_t)}{\partial t}dt \\ & + e^{-rt}\frac{\partial V(t,S_t)}{\partial S_t} S_t^-r dt + e^{-rt}\frac{\partial V(t,S_t)}{\partial S_t} S_t^-\sigma dW_t \\ & + e^{-rt}\frac{\partial V(t,S_t)}{\partial S_t} S_t^-\int\limits_{\mathbb{R}} (e^x - 1) J_X (dt \times dx) \\ & - e^{-rt}\frac{\partial V(t,S_t)}{\partial S_t} S_t^-\int\limits_{\mathbb{R}} (e^x - 1) \nu (dt) dx \\ & + e^{-rt}\frac{\partial^2 V(t,S_t)}{\partial S_t^{\,\,2}} \frac{\sigma^2}{2} dt + e^{-rt} \int\limits_{\mathbb{R}} [ V(t,S_{t^-}e^x ) - V(t,S_{t^-}) ] J_X (dt \times dx) \\ & - e^{-rt} \frac{\partial V(t,S_t)}{\partial S_t}\int\limits_{\mathbb{R}} S_{t^-} (e^x - 1) J_X (dt \times dx) \\ d( e^{-rt} V(t,S_t)) = & -r e^{-rt}V(t,S_t) dt + e^{-rt} \frac{\partial V(t,S_t)}{\partial t}dt \\ & + e^{-rt}\frac{\partial V(t,S_t)}{\partial S_t} S_t^-r dt + e^{-rt}\frac{\partial V(t,S_t)}{\partial S_t} S_t^-\sigma dW_t \\ & - e^{-rt}\frac{\partial V(t,S_t)}{\partial S_t} S_t^-\int\limits_{\mathbb{R}} (e^x - 1) \nu (dt) dx + e^{-rt}\frac{\partial^2 V(t,S_t)}{\partial S_t^{\,\,2}} \frac{\sigma^2}{2} dt \\ & + e^{-rt} \int\limits_{\mathbb{R}} [ V(t,S_{t^-}e^x ) - V(t,S_{t^-}) ] \tilde{J}_X (dt \times dx) \\ & + e^{-rt} \int\limits_{\mathbb{R}} [ V(t,S_{t^-}e^x ) - V(t,S_{t^-}) ] \nu (dx) dt \end{align*}

We can split the formula with a martingale part and a finite variation part. We have:

∀t∈R,dAt=e−rt [ −rV(t,St)dt+∂V(t,St)∂tdt+∂V(t,St)∂StSt−rdt+∂2V(t,St)∂St  2σ22dt+∫R[V(t,St−ex)−V(t,St−)]ν(dx)dt−∂V(t,St)∂St∫RSt−(ex−1)ν(dx)dt ]∀t∈R,dMt=e−rt [ ∂V(t,St)∂StSt−σdWt+∫R[V(t,St−ex)−V(t,St−)]J~X(dt×dx) ] \begin{align*} \forall t \in \mathbb{R}, d A_t = & e^{-rt} \, [ \, -r V(t,S_t) dt + \frac{\partial V(t,S_t)}{\partial t}dt + \frac{\partial V(t,S_t)}{\partial S_t} S_{t^-} r dt + \frac{\partial^2 V(t,S_t)}{\partial S_t^{\,\,2}} \frac{\sigma^2}{2} dt \\ & + \int\limits_{\mathbb{R}} [ V(t,S_{t^-}e^x ) - V(t,S_{t^-}) ] \nu (dx) dt - \frac{\partial V(t,S_t)}{\partial S_t}\int\limits_{\mathbb{R}} S_{t^-} (e^x - 1) \nu (dx) dt \, ] \\ \forall t \in \mathbb{R}, d M_t = & e^{-rt} \, [ \, \frac{\partial V(t,S_t)}{\partial S_t} S_{t^-} \sigma dW_t + \int\limits_{\mathbb{R}} [ V(t,S_{t^-}e^x ) - V(t,S_{t^-}) ] \tilde{J}_X (dt \times dx) \, ] \end{align*}

with ∀t∈R \forall t \in \mathbb{R} , AtA_t the finite variation part and ∀t∈R \forall t \in \mathbb{R} , MtM_t the martingale part.

We have to cancel the finite variation part. The partial integro-differential equation is therefore, ∀t∈R \forall t \in \mathbb{R} ,

0=e−rt [ −rV(t,St)dt+∂V(t,St)∂tdt+∂V(t,St)∂StSt−rdt+∂2V(t,St)∂St  2σ22dt+∫R[V(t,St−ex)−V(t,St−)]ν(dx)dt−∂V(t,St)∂St∫RSt−(ex−1)ν(dx)dt ]rV(t,St)=∂V(t,St)∂t+∂V(t,St)∂StSt−r+∂2V(t,St)∂St  2σ22+∫R [ V(t,St−ex)−V(t,St−)−∂V(t,St)∂StSt−(ex−1) ] ν(dx)rV(t,St)=∂V(t,St)∂t+∂V(t,St)∂StSt− [ r+∫R St−(ex−1) ν(dx) ] +∂2V(t,St)∂St  2σ22+∫RV(t,St−ex)−V(t,St−) ν(dx) \begin{align*} 0 = & e^{-rt} \, [ \, -r V(t,S_t) dt + \frac{\partial V(t,S_t)}{\partial t}dt + \frac{\partial V(t,S_t)}{\partial S_t} S_{t^-} r dt + \frac{\partial^2 V(t,S_t)}{\partial S_t^{\,\,2}} \frac{\sigma^2}{2} dt \\ & + \int\limits_{\mathbb{R}} [ V(t,S_{t^-}e^x ) - V(t,S_{t^-}) ] \nu (dx) dt - \frac{\partial V(t,S_t)}{\partial S_t}\int\limits_{\mathbb{R}} S_{t^-} (e^x - 1) \nu (dx) dt \, ] \\ r V(t,S_t) = & \frac{\partial V(t,S_t)}{\partial t} + \frac{\partial V(t,S_t)}{\partial S_t} S_{t^-} r + \frac{\partial^2 V(t,S_t)}{\partial S_t^{\,\,2}} \frac{\sigma^2}{2} \\ & + \int\limits_{\mathbb{R}} \, [ \, V(t,S_{t^-}e^x ) - V(t,S_{t^-}) - \frac{\partial V(t,S_t)}{\partial S_t} S_{t^-} (e^x - 1) \, ] \, \nu (dx) \\ r V(t,S_t) = & \frac{\partial V(t,S_t)}{\partial t} + \frac{\partial V(t,S_t)}{\partial S_t} S_{t^-} \, [ \, r + \int\limits_{\mathbb{R}} \, S_{t^-} (e^x - 1) \, \nu (dx) \, ] \, \\ & + \frac{\partial^2 V(t,S_t)}{\partial S_t^{\,\,2}} \frac{\sigma^2}{2} + \int\limits_{\mathbb{R}} V(t,S_{t^-}e^x ) - V(t,S_{t^-}) \, \nu (dx) \end{align*}

Remark: If the Lévy measure is always equals to 0 then the PIDE is in fact the well known PDE applying to continuous Lévy processes.

For dStSt=rdt+σdWt \frac{d S_t}{S_t} = r dt + \sigma d W_t , we have the following valuation PDE:

∀t∈R,rV(t,St)=∂V(t,St)∂t+∂V(t,St)∂StStr+∂2V(t,St)∂St  2σ22 \forall t \in \mathbb{R}, r V(t,S_t) = \displaystyle \frac{\partial V(t,S_t)}{\partial t} + \frac{\partial V(t,S_t)}{\partial S_t} S_t r + \frac{\partial^2 V(t,S_t)}{\partial S_t^{\,\,2}} \frac{\sigma^2}{2}

Jump diffusion models

We decide to focus ourselves on a subset of jump diffusion models. We expose the advantage of introducing jumps in the underlyings dynamic. We find the dynamic and the risk neutral probability. The reader can expand/collapse for more information

Why Jump diffusion models?

Empirically we observe the following characteristic for the underlying returns:

  • They don’t follow a log normal law. We observe heavy-tailed distribution and the kurtosis is not null
  • They are not continuous
  • We observe the volatility clustering phenomenon

Underlyings dynamic (for finite activity)

Let WW be a brownian motion and JJ a composed Poisson process. We note the filtration (Ft)=(σ(Ws)s≤t)∨(σ(Js)s≤t) (\mathcal{F}_t) = ( \sigma(W_s)_{s \le t} ) \lor ( \sigma(J_s)_{s \le t} ) with ∨ \lor defined by F1∨F2 \mathcal{F}_1 \lor \mathcal{F}_2 is the smallest filtration including F1⋃F2 \mathcal{F}_1 \bigcup \mathcal{F}_2 .

The dynamic in the jump model framework is split in two parts: a continuous diffusion part and a jump part.

the dynamic under the risk neutral probability is:

dStSt−=μ(t−,St−)dt+σ(t−,St−)dWt⏟continuous  part+dJt⏟jump  partwith Jt=∑i=1Nt(Ai−1) \displaystyle \frac{d S_t}{S_{t^{-}}} =\underbrace{\mu (t^{-},S_{t^{-}}) dt + \sigma ({t^{-}},S_{t^{-}}) d W_t}_{continuous \; part} + \underbrace{d J_t}_{jump \; part} \text{with } J_t = \displaystyle \sum_{ i = 1}^{N_t} ( A_i - 1 )

let’s note dStSt−=dXt \displaystyle \frac{d S_t}{S_{t^{-}}} = d X_t . We can compute both parts:

  • dXtc=μ(t−,St−)dt+σ(t−,St−)dWt d X_t^c = \mu ({t^{-}},S_{t^{-}}) dt + \sigma ({t^-},S_{t^-}) d W_t so Xtc−X0c=∫0tμ(s−,Ss−)ds+∫0tσ(s−,Ss−)dWs X_t^c - X_0^c = \displaystyle \int\limits_0^t \mu ({s^{-}},S_{s^{-}}) ds + \displaystyle \int\limits_0^t \sigma ({s^{-}},S_{s^{-}}) d W_s
  • d[X]tc=σ(t−,St−)2dt d [X]_t^c = \sigma({t^{-}},S_{t^{-}})^2 d t so [X]tc=∫0tσ(s−,Ss−)2ds [X]_t^c = \displaystyle \int\limits_0^t \sigma ({s^{-}},S_{s^{-}})^2 ds
  • ∏s≤t(1+△Xs)=∏i=1NtAi \displaystyle \prod_{ s \le t } ( 1 + \triangle X_s ) = \displaystyle \prod_{ i = 1 }^{N_t} A_i

The underlying price is:

St=S0exp⁡[∫0t(μ(s−,Ss−)−σ(s−,Ss−)22) ds+∫0tσ(s−,Ss−) dWs]∏i=1NtAi S_t = \displaystyle S_0 \exp \left[ \int\limits_{0}^t ( \mu ({s^-},S_{s^-}) - \frac{\sigma ({s^-},S_{s^-})^2}{2} ) \, \mathrm ds + \int\limits_{0}^t \sigma ({s^-},S_{s^-}) \, \mathrm d W_s \right] \displaystyle \prod_{i=1}^{N_t} A_i

Risk neutral probabilities

Let’s note μa=E[A1]=E[ANt] \mu_a = \mathbb{E} [ A_1] = \mathbb{E} [ A_{N_t}], the mean of the jump size. First of all, we have to prove that (Jt−μa∫0tλ(s)ds)t ( J_t - \mu_a \displaystyle \int\limits_0^t \lambda (s) ds )_{t} is a (Ft \mathcal{F}_t )-martingale.

Let’s note M(t)=Jt−μa∫0tλ(s)ds M(t) = J_t - \mu_a \displaystyle \int\limits_0^t \lambda (s) ds .

We have E[Jt]=μa∫0tλ(s)ds\mathbb{E} [ J_t ] = \mu_a \displaystyle \int\limits_0^t \lambda (s) ds .

E[M(t) ∣ Fs]=E[Jt−μa∫0tλ(s)ds ∣ Fs]=E[Jt−Js−μa∫stλ(u)du ∣ Fs]+Js−μa∫0sλ(u)du=E[∑k=1+NsNtAi ∣ Fs]−μa∫stλ(u)du⏟=0+Js−μa∫0sλ(u)du=Js−μa∫0sλ(u)du=M(s) \begin{align*} \mathbb{E} [ M(t) \, | \, \mathcal{F}_s ] & = \mathbb{E} [ J_t - \mu_a \displaystyle \int\limits_0^t \lambda (s) ds \, | \, \mathcal{F}_s ] = \mathbb{E} [ J_t - J_s - \mu_a \displaystyle \int\limits_s^t \lambda (u) du \, | \, \mathcal{F}_s ] + J_s - \mu_a \displaystyle \int\limits_0^s \lambda (u) du \\ & = \underbrace{\mathbb{E} [ \displaystyle \sum_{k= 1+N_s}^{N_t} A_i \, | \, \mathcal{F}_s ] - \mu_a \displaystyle \int\limits_s^t \lambda (u) du }_{=0} + J_s - \mu_a \displaystyle \int\limits_0^s \lambda (u) du = J_s - \mu_a \displaystyle \int\limits_0^s \lambda (u) du = M(s) \end{align*}

Because E[M(t) ∣ Fs]=Js−μa∫0sλ(u)du=M(s) \mathbb{E} [ M(t) \, | \, \mathcal{F}_s ] = J_s - \mu_a \displaystyle \int\limits_0^s \lambda (u) du = M(s) with dStSt−=μ(t−,St−)dt+σ(t−,St−)dWt+dJt \frac{d S_t}{S_{t^{-}}} = \mu (t^{-},S_{t^{-}}) dt + \sigma ({t^{-}},S_{t^{-}}) d W_t + d J_t .

The discount factor is noted DtD_t and is equal to: Dt=e−∫0tr(s) dsD_t = e^{ - \int\limits_{0}^t r(s) \, \mathrm ds}

d(DtSt)=−r(t)DtStdt+DtdStDtSt=∫0tDs−Ss−[(μ(s−,Ss−)−r(s))ds+σ(s−,SS−)dWs+dJs]+D0S0DtSt=∫0tDs−Ss−[μ(s−,Ss−)+λ(s−)(μa−1)−r(s−)]ds+∫0tDs−Ss−σ(s−,Ss−)dWs+∫0tDs−Ss−dMs+D0S0 \begin{align*} d ( D_t S_t ) & = - r(t) D_t S_t dt + D_t d S_t \\ D_t S_t & = \displaystyle \int\limits_0^t D_{s^-} S_{{s^-}} [ ( \mu (s^{-},S_{s^{-}})- r(s) ) ds + \sigma ({s^{-}},S_{S^{-}}) d W_s + d J_s ] + D_0 S_0 \\ D_t S_t & = \displaystyle \int\limits_0^t D_{s^-} S_{{s^-}} [ \mu (s^{-},S_{s^{-}}) + \lambda(s^-) (\mu_a - 1 ) - r(s^-) ] ds \\ & + \displaystyle \int\limits_0^t D_{s^-} S_{{s^-}} \sigma ({s^{-}},S_{s^{-}}) d W_s \displaystyle + \int\limits_0^t D_{s^-} S_{{s^-}} d M_s + D_0 S_0 \\ \end{align*}

Therefore

DtSt−∫0tDs−Ss−σ(s−,Ss−)dWs−∫0tDs−Ss−dMs=∫0tDs−Ss−[μ(s−,Ss−)+λ(s−)(μa−1)−r(s−)]ds+D0S0 \begin{align*} & D_t S_t - \displaystyle \int\limits_0^t D_{s^-} S_{{s^-}} \sigma ({s^{-}},S_{s^{-}}) d W_s - \displaystyle \int\limits_0^t D_{s^-} S_{{s^-}} d M_s \\ & = \displaystyle \int\limits_0^t D_{s^-} S_{{s^-}} [ \mu (s^{-},S_{s^{-}}) + \lambda(s^-) (\mu_a - 1 ) - r(s^-) ] ds + D_0 S_0 \end{align*}

On the left side is a linear combination of martingale (an integral again a martingale is still a martingale).

Therefore, the right side is also a martingale. We have:

∫0⋅(μ(u−,Su−)−ru−+λu−[μa−1]  ) du \int\limits_{0}^{\cdot} ( \mu ({u^-},S_{u^-}) - r_{u^-} + \lambda_{u^-} [ \mu_a - 1] \; )\, \mathrm du

which is a martingale so:

∀t∈R+,  μ(t,St)−rt+λt[μa−1]=0⇔  ∀t  μ(t,St)=rt−λt[μa−1] \forall t \in \mathbb{R}^+ ,\; \mu (t,S_t) - r_t + \lambda_{t} [ \mu_a - 1] = 0 \Leftrightarrow \; \forall t \; \mu (t,S_t) = r_t - \lambda_{t} [ \mu_a - 1]

Under the risk neutral probability, the underlying price is:

St=S0exp⁡[∫0t(ru−−λu−[μa−1]−σ(u−,Su−)22) du+∫0tσ(u−,Su−) dWu]∏i=1NtAi S_t = S_0 \displaystyle \exp \left[ \int\limits_{0}^t \left( r_{u^-} - \lambda_{u^-} [ \mu_a - 1] - \frac{\sigma ({u^-},S_{u^-})^2}{2} \right) \, \mathrm du + \int\limits_{0}^t \sigma ({u^-},S_{u^-}) \, \mathrm d W_u \right] \displaystyle \prod_{i=1}^{N_t} A_i

Merton jump diffusion model

A model having this features is the Merton model. We expose the dynamic of the underlyings and the european call price formula. We also investigate the calibration procedure in this model.

The dynamic in this model is the following one:

dStSt−=(rt−−λt−[μa−1])dt+σt−dWt+dJt \frac{d S_t}{S_{t^{-}}} = ( r_{t^-} - \lambda_{t^-} [ \mu_a - 1] ) dt + \sigma_{t^-} d W_t + d J_t

with

  • the interest rate rr and the volatility σ\sigma are supposed to be deterministic functions
  • ∀μz,σz∈R2Ai=eμz+σzZi \forall \mu_z, \sigma_z \in \mathbb{R}^2 A_i = e^{\mu_z + \sigma_z Zi}
  • ∀n,(Z1,…,Zn)∼N(0,In) \forall n, (Z_1, \ldots ,Z_n ) \sim \mathcal{N}(0, I_n)
  • ∀i∈N∗,∀t∈R+,Ai⊥ ⁣ ⁣ ⁣⊥Nt \forall i \in \mathbb{N}^*, \forall t \in \mathbb{R}^+, A_i \perp\!\!\!\perp N_t
  • μa=exp(μz+σz22) \mu_a = exp( \mu_z + \frac{\sigma_z^2}{2} )

The underlying price is

St=S0exp⁡(∫0t(rs−−λs−[μa−1]−σs−22) ds+∫0tσs− dWs)∏i=1NtAi=S0exp⁡( ∫0t(rs−−λs−[μa−1]−σs−22) ds+∫0tσs− dWs+∑i=1Nt[μz+σz  Zi] ) \begin{align*} S_t & = \displaystyle S_0 \exp \left( \int\limits_{0}^t ( r_{s^-} - \lambda_{s^-} [ \mu_a - 1] - \frac{\sigma_{s^-}^2}{2} ) \, \mathrm ds + \int\limits_{0}^t \sigma_{s^-} \, \mathrm d W_s \right) \displaystyle \prod_{i=1}^{N_t} A_i \\ & = S_0 \exp \left( \, \int\limits_{0}^t ( r_{s^-} - \lambda_{s^-} [ \mu_a - 1] - \frac{\sigma_{s^-}^2}{2} ) \, \mathrm ds + \int\limits_{0}^t \sigma_{s^-} \, \mathrm d W_s + \displaystyle \sum_{i=1}^{N_t} [ \mu_z + \sigma_z \; Z_i ] \, \right) \end{align*}

European call price

We expose in this section three different methods to compute a call price in the Merton model:

  • Monte-Carlo
  • PIDE
  • Series expansion

For Monte-Carlo and PIDE methods, we refer the reader to the book of R. Cont and P. Tankov: “Financial Modelling with Jump Processes”. The reader can expand/collapse for more information on the series expansion method

Let’s compute the call price under the risk neutral probability previously exposed. A same computation can be apply for Put price.

The steps are the followings:

  • Apply the law of iterated expectations using the Poisson process conditioning
  • Compute the new log normal parameters
  • Compute the expected value of a Black-Scholes call as a function of the Poisson law

Let’s compute the call price:

E[e−∫strudu  (St−K)+ ∣ Fs]=E[  E[  e−∫strudu  (St−K)+∣Nt=n  ]   ∣ Fs] \mathbb{E} \left[ e^{ - \int\limits_{s}^t r_u \mathrm du } \; ( S_t - K )_{+} \, | \, \mathcal{F}_s \right] = \mathbb{E} \left[ \; \mathbb{E} [ \; e^{ - \int\limits_{s}^t r_u \mathrm du } \; ( S_t - K )_{+} | N_t =n \; ] \; \, | \, \mathcal{F}_s \right]

Let’s identify the law of St S_t conditional to the event {Nt=n N_t = n }:

St∣Nt,Fs=LSsexp⁡(  ∫st(ru−−λu−[μa−1]−σu−22) du+∫stσu− dWu+nμz+nσz  Z1  ) S_t | N_t , \mathcal{F}_s \overset{\mathcal{L}}{=} S_s \exp \left( \; \int\limits_{s}^t ( r_{u^-} - \lambda_{u^-} [ \mu_a - 1] - \frac{\sigma_{u^-}^2}{2} ) \, \mathrm du + \int\limits_{s}^t \sigma_{u^-} \, \mathrm d W_u + n \mu_z + \sqrt{n} \sigma_z \; Z_1 \; \right)

Moreover, we know that ∫stσu dWu∼N(0,∫stσu2 du) \int\limits_{s}^t \sigma_u \, \mathrm d W_u \sim N ( 0 , \int\limits_{s}^t \sigma_u^2 \, \mathrm du ) .

Let’s write the integral against the brownian motion using law equalities:

∫stσu− dWu+nσzL=L(∫stσu−2 du+nσz2)  L=L(∫stσu−2 du+nσz2)  (Wt−Ws)t−s=L1t−s(∫stσu−2 du+nσz2)(Wt−Ws) \begin{align*} \int\limits_{s}^t \sigma_{u^-} \, \mathrm d W_u + \sqrt{n} \sigma_z L &\overset{\mathcal{L}}{=} \sqrt{ ( \int\limits_{s}^t \sigma_{u^-}^2 \, \mathrm du + n \sigma_z^2 } ) \; L \overset{\mathcal{L}}{=} \sqrt{( \int\limits_{s}^t \sigma_{u^-}^2 \, \mathrm du + n \sigma_z^2 )} \; \frac{( W_t - W_s )}{ \sqrt{t-s} }\\ &\overset{\mathcal{L}}{=} \sqrt{ \frac{1}{ t - s } \left( \int\limits_{s}^t \sigma_{u^-}^2 \, \mathrm du + n \sigma_z^2 \right) } ( W_t - W_s ) \end{align*}

with LL which follow a standard normal distribution. We have in law:

St∣Nt,Fs=LSsexp⁡(  ∫st(ru−−λu−[μa−1]−σu−22) du+nμz+∫stσu−2 dut−s+nσz2t−s  (Wt−Ws)  ) S_t | N_t , \mathcal{F}_s \overset{\mathcal{L}}{=} S_s \exp \left( \; \int\limits_{s}^t ( r_{u^-} - \lambda_{u^-} [ \mu_a - 1] - \frac{\sigma_{u^-}^2}{2} ) \, \mathrm du + n \mu_z + \sqrt{ \frac{\int\limits_{s}^t \sigma_{u^-}^2 \, \mathrm du }{ t - s } + \frac{n \sigma_z^2}{ t - s } } \; ( W_t - W_s ) \; \right) St∣Nt,Fs=LSsexp⁡(nμz−∫stλu−du[μa−1]+nσz22)exp⁡(  ∫stru−−σu−22 du−nσz22+∫stσu−2 dut−s+nσz2t−s  (Wt−Ws)  ) \begin{align*} S_t | N_t , \mathcal{F}_s \overset{\mathcal{L}}{=} S_s & \exp \left( n \mu_z - \int\limits_{s}^t \lambda_{u^-} du [ \mu_a - 1] + \frac{n \sigma_z^2}{ 2 } \right) \\ & \exp \left( \; \int\limits_{s}^t r_{u^-} - \frac{\sigma_{u^-}^2}{2} \, \mathrm du - \frac{n \sigma_z^2}{ 2 } + \sqrt{ \frac{\int\limits_{s}^t \sigma_{u^-}^2 \, \mathrm du }{ t - s } + \frac{n \sigma_z^2}{ t - s } } \; ( W_t - W_s ) \; \right) \end{align*}

We can therefore note Sˉs(n) \bar S_s (n) the new Ss S_s and σˉ(n) \bar \sigma (n) the new σt \sigma_t :

Sˉs(n)=Ssexp⁡(∫stλu−du[1−μa]+nμz+nσz22)σˉ(n)=nσz2t−s+∫stσs−2dst−s \begin{align*} \bar S_s (n) & = S_s \exp \left( \int\limits_{s}^t \lambda_{u^-} du [1 -\mu_a] + n \mu_z + \frac{n \sigma_z^2}{ 2 } \right) \\ \bar \sigma (n) & = \sqrt{ \frac{n \sigma_z^2 }{t-s} + \frac{ \int\limits_{s}^t \sigma_{s^-}^2 ds}{t-s} } \end{align*}

We have:

St∣Nt,Fs=LSˉs(Nt)exp⁡(  ∫stru− du−σˉ(Nt)22(t−s)+σˉ(Nt)  (Wt−Ws)  ) S_t | N_t , \mathcal{F}_s \overset{\mathcal{L}}{=} \bar S_s (N_t) \exp \left( \; \int\limits_{s}^t r_{u^-} \, \mathrm du - \frac{ \bar \sigma (N_t)^2}{2} (t-s) + \bar \sigma (N_t) \; (W_t - W_s) \; \right)

Let’s back to the initial conditional expectation:

E[  E[e−∫strudu  (St−K)+∣Nt=n  ]  ∣Fs]=E[CBS(Ssˉ(k),K,r,t,σˉ(k))∣Fs] \mathbb{E} \left[ \; \mathbb{E} \left[ e^{- \int\limits_{s}^t r_u \mathrm du } \; ( S_t - K )_{+} | N_t =n \; \right] \; | \mathcal{F}_s \right] = \mathbb{E} \left[ C_{BS}( \bar{S_s}(k),K,r,t,\bar{\sigma}(k)) | \mathcal{F}_s \right]

Under the Black-Scholes model, we recall the european call closed form:

CBS(Ss,K,r,T,σ)=SsN(d1)−Kexp(−rT)N(d2)d1=1σT[ln(SsK)+(r+σ22)T]d2=d1−σT \begin{align*} C_{BS} & ( S_s, K, r, T,\sigma ) = S_s N(d1) - K exp(-r T) N(d2) \\ d1 & = \frac{1}{\sigma \sqrt{T}} [ ln( \frac{S_s}{K} ) + ( r + \frac{\sigma^2}{2} ) T ] \\ d2 & = d1 - \sigma \sqrt{T} \end{align*}

We therefore have to compute the CBS C_{BS} with the new parameters previously found.

The Merton call price is given by:

Call(s,T,S,K)=∑k=0∞P(Nt=k)×CBS(Ssˉ(Nt),K,r,t,σˉ(Nt)) Call (s,T,S,K) = \displaystyle \sum_{k = 0}^{\infty} \mathbb{P} ( N_t = k ) \times C_{BS}( \bar{S_s}(N_t),K,r,t,\bar{\sigma}(N_t) )

Error control in call pricing

We have to find a stopping criteria for the previous series expansion, we want to identify kopt k_{opt} such that

Err(kopt)=∣1S0(∑k=0∞f(k)−∑k=0koptf(k))∣<ϵf(k)=P(Nt=k)×T(k)T(k)=CBS(Ssˉ(k),K,r,t,σˉ(k)) \begin{align*} Err (k_{opt}) = & |\frac{1}{S_0} ( \displaystyle \sum_{k = 0}^{\infty} f(k) - \displaystyle \sum_{k = 0}^{k_{opt}} f(k) ) | < \epsilon \\ f(k) = & \mathbb{P} ( N_t = k ) \times T(k) \\ T(k) = & C_{BS}( \bar{S_s}(k),K,r,t,\bar{\sigma}(k)) \end{align*}

Let’s suppose λ \lambda constant for the purpose of clarity

S0×Err(k)=∣∑k=0∞e−λt(λt)kk!T(k)−∑k=0kopte−λt(λt)kk!T(k)∣=∣∑k=kopt+1∞e−λt(λt)kk!T(k)∣=∣∑k=kopt+1∞P(Nt=k)T(k)∣ \begin{align*} S_0 \times Err(k) & = | \displaystyle \sum_{k = 0}^{\infty} {e^{-\lambda t}} {\frac{(\lambda t)^k}{k!} } T(k) - \displaystyle \sum_{k = 0}^{k_{opt}} {e^{-\lambda t}} {\frac{(\lambda t)^k}{k!} } T(k) | = | \displaystyle \sum_{k = k_{opt}+1 }^{\infty} {e^{-\lambda t}} {\frac{(\lambda t)^k}{k!} } T(k)| \\ & = | \displaystyle \sum_{k = k_{opt}+1 }^{\infty} \mathbb{P} ( N_t = k ) T(k)| \end{align*}

Using the non arbitrary conditions, we can write T(k)≤S0ˉ T(k) \le \bar{S_0} .

Moreover, all the terms of the sum are positive, so:

S0×Err(k)≤∑k=kopt+1∞e−λt(λt)kk!Sˉ0=e−λt∑k=kopt+1∞(λt)kk!S0exp⁡(λt[1−μa]+kμz+kσz22) S_0 \times Err(k) \le \displaystyle \sum_{k = k_{opt}+1 }^{\infty} {e^{-\lambda t}} {\frac{(\lambda t)^k}{k!} } \bar{S}_0 = {e^{-\lambda t}} \displaystyle \sum_{k = k_{opt}+1 }^{\infty} {\frac{(\lambda t)^k}{k!} } S_0 \exp ( \lambda t [1 -\mu_a] + k \mu_z + \frac{k \sigma_z^2}{ 2 } )

=e−λtμa∑k=kopt+1∞(λt)kk!S0exp⁡(kμz+kσz22)=S0e−λtμa∑k=kopt+1∞(λt)kk!μak = {e^{ - \lambda t \mu_a } } \displaystyle \sum_{k = k_{opt}+1 }^{\infty} {\frac{(\lambda t)^k}{k!} } S_0 \exp ( k \mu_z + \frac{k \sigma_z^2}{ 2 } ) = S_0 {e^{- \lambda t \mu_a } } \displaystyle \sum_{k = k_{opt}+1 }^{\infty} {\frac{(\lambda t)^k}{k!} } \mu_a^k

Let’s note y=λtμa y = \lambda t \mu_a ,

S0×Err(kopt)≤S0e−y∑k=kopt+1∞ykk!=S0e−y[ey−∑k=0koptykk!] S_0 \times Err ( k_{opt} ) \le S_0 {e^{-y} } \displaystyle \sum_{k = k_{opt}+1 }^{\infty} {\frac{y^k}{k!} } = S_0 {e^{-y} } [ e^{y} - \displaystyle \sum_{k = 0 }^{k_{opt}} {\frac{y^k}{k!} } ]

Err(kopt)≤e−y[ey−∑k=0koptykk!] Err ( k_{opt} ) \le {e^{-y} } [ e^{y} - \displaystyle \sum_{k = 0 }^{k_{opt}} {\frac{y^k}{k!} } ]

We simply have to test if ey−∑k=0kopt(y)kk! e^{y} - \displaystyle \sum_{k = 0}^{k_{opt}}{\frac{(y)^k}{k!} } is lower than a chosen tolerance.

Numerically, this sum can be directly computed. We have to accumulate the term inside the price computation.

Greeks computation

Greeks can be computed using a closed form by applying the chain rule. The reader can expand/collapse for more details on greeks.

δ=∂Call∂S=∑k=0∞e−λt(λt)kk!×δBS(k)×exp⁡(∫0tλsds[1−μa]+kμz+kσz22)γ=∂2Call∂S2=∑k=0∞e−λt(λt)kk!×γBS(k)×exp⁡(2[∫0tλsds[1−μa]+kμz+kσz22])vega=∂Call∂σ=∑k=0∞e−λt(λt)kk!×vegaBS(k)×σttnσz2+∫0tσt2∂Call∂K=∑k=0∞e−λt(λt)kk!×∂CBS∂K(k)∂2Call∂K2=∑k=0∞e−λt(λt)kk!×∂2CBS∂K2(k) \begin{align*} \delta = \frac{\partial Call }{\partial S } = & \displaystyle \sum_{k = 0}^{\infty} {e^{-\lambda t}} {\frac{(\lambda t)^k}{k!} } \times \delta^{BS}(k) \times \exp \left( \int\limits_{0}^t \lambda_s ds [1 -\mu_a] + k \mu_z + \frac{k \sigma_z^2}{ 2 } \right) \\ \gamma = \frac{\partial^2 Call }{\partial S^2 } = & \displaystyle \sum_{k = 0}^{\infty} {e^{-\lambda t}} {\frac{(\lambda t)^k}{k!} } \times \gamma^{BS}(k) \times \exp \left( 2 [ \int\limits_{0}^t \lambda_s ds [1 -\mu_a] + k \mu_z + \frac{k \sigma_z^2}{ 2 } ] \right) \\ vega = \frac{\partial Call }{\partial \sigma } = & \displaystyle \sum_{k = 0}^{\infty} {e^{-\lambda t}} {\frac{(\lambda t)^k}{k!} } \times vega^{BS}(k) \times \frac{\sigma_t \sqrt{t}} {n \sigma_z^2 + \int\limits_{0}^t \sigma_t^2 } \\ \frac{\partial Call }{\partial K } & = \displaystyle \sum_{k = 0}^{\infty} {e^{-\lambda t}} {\frac{(\lambda t)^k}{k!} } \times \frac{\partial C_{BS}}{\partial K }(k) \\ \frac{\partial^2 Call }{\partial K^2 } & = \displaystyle \sum_{k = 0}^{\infty} {e^{-\lambda t}} {\frac{(\lambda t)^k}{k!} } \times \frac{\partial^2 C_{BS} }{\partial K^2 }(k) \end{align*}

Generic calibration procedure

Let introduce the classical calibration procedure.

arg⁡p∈Pmin⁡∑i∈Iwi(σimod(p)−σimkt)2 \underset{p \in \mathcal{P}}{\arg} \min { \displaystyle \sum_{i \in I} w_i { ( \sigma_i^{mod}(p) - \sigma_i^{mkt} )^2 } }

with

  • P \mathcal{P} set of model parameters
  • II set of indexes related to the quoted options
  • (wi)i∈I {(w_i})_{i \in I} a weight family
  • σimkt \sigma_i^{mkt} implied volatility for the ithi^{th} option
  • σimod(p) \sigma_i^{mod} (p) equivalent Black-Scholes volatility using pp parameters
  • ViBSV_i^{BS}: BS price for the ithi^{th} option.
  • VegaBSi=∂VBS∂σ VegaBS_i = \displaystyle \frac{ \partial {V^{BS}} }{ \partial \sigma }

By definition,

Vimod=ViBS(σimod(p))Vimkt=ViBS(σimkt) \begin{align*} & V_i^{mod} = V_i^{BS} (\sigma_i^{mod}(p)) \\ & V_i^{mkt} = V_i^{BS} (\sigma_i^{mkt}) \end{align*}

Using Taylor series, we have:

ViBS(σimod(p))=ViBS(σimkt)+(σimod(p)−σimkt)∂ViBS∂σ(σimkt)+O(σimod(p)−σimkt)2 V_i^{BS} (\sigma_i^{mod}(p)) = V_i^{BS} (\sigma_i^{mkt}) + ( \sigma_i^{mod}(p) - \sigma_i^{mkt} ) \displaystyle \frac{\partial V_i^{BS} }{\partial \sigma } (\sigma_i^{mkt}) + O ( \sigma_i^{mod}(p) - \sigma_i^{mkt} )^2 σimod(p)−σimkt=ViBS(σimod(p))−ViBS(σimkt)∂ViBS∂σ(σimkt)+O(σimod(p)−σimkt)2=ViBS(σimod)(p)−ViBS(σimkt)VegaBSi(σimkt)+O(σimod(p)−σimkt)2 \begin{align*} \sigma_i^{mod}(p) - \sigma_i^{mkt} & = \displaystyle \frac{ V_i^{BS} (\sigma_i^{mod}(p)) - V_i^{BS} (\sigma_i^{mkt}) }{ \frac{\partial V_i^{BS} }{\partial \sigma } (\sigma_i^{mkt}) } + O ( \sigma_i^{mod}(p) - \sigma_i^{mkt} )^2 \\ & = \displaystyle \frac{ V_i^{BS} (\sigma_i^{mod})(p) - V_i^{BS} (\sigma_i^{mkt}) }{ VegaBS_i (\sigma_i^{mkt}) } + O ( \sigma_i^{mod}(p) - \sigma_i^{mkt} )^2 \end{align*} arg⁡p∈Pmin⁡∑i∈Iwi[VBS(σimod(p))−VBS(σimkt)VegaBSi(σimkt)]2 \underset{p \in \mathcal{P}}{\arg} \min \displaystyle \sum_{i \in I} w_i \left[ { \frac{ V^{BS} (\sigma_i^{mod}(p)) - V^{BS} (\sigma_i^{mkt}) }{ VegaBS_i (\sigma_i^{mkt}) } } \right]^2

Parameters impact

To expose the parameters influence, we expose the implied volatility surface in function of the model parameters. Don’t hesitate to expand this section

Let’s have the following parameters:

  • s0 = 100 and r = 0.02
  • lambda = 0.2
  • sigma = 0.3
  • mu_z = -0.3
  • sigma_z = 0.1
This will display an animated GIF

Bates Model

This model merges the benefits of Merton and Heston models.

The underlying dynamic is based on the Black-Scholes one with two additional features:

  • Jump part: control the smile for short term smile (< 1 an)
  • Stochastic volatility part, follow a Cox Ingersoll Ross (CIR) process: control the smile for higher maturities (> 1 an)

General dynamic

Let WsW^s and WvW^v two brownian motions with correlation ρ∈[−1,1]\rho \in \mathopen{[}-1,1\mathclose{]}.

Let JJ be a composed Poisson process.

Let have the filtration (Ft)=(σ(Wus)u≤t)∨(σ(Wuv)u≤t)∨(σ(Ju)u≤t) (\mathcal{F}_t) = ( \sigma(W_u^s)_{u \le t} ) \lor ( \sigma(W_u^v)_{u \le t} ) \lor ( \sigma(J_u)_{u \le t} )

The underlying dynamic is the following one:

dStSt−=(rt+λ(1−μa))dt+Vt−dWts+dJtdVt=ξ(η−Vt)dt+θVtdWtv \begin{align*} \frac{d S_t}{S_{t^{-}}} & = ( r_t + \lambda ( 1 - \mu_a) ) dt + \sqrt{V_{t^{-}}} d W_t^s+ d J_t \\ dV_{t} & = \xi (\eta - V_{t}) dt + \theta \sqrt{V_{t}} d W_t^v \end{align*}

with

  • Jt=∑i=1Nt(Ai−1) J_t = \displaystyle \sum_{ i = 1}^{N_t} ( A_i - 1 ) with ∀i∈N∗,∀t∈R+,Ai⊥ ⁣ ⁣ ⁣⊥Nt \forall i \in \mathbb{N}^*, \forall t \in \mathbb{R}^+, A_i \perp\!\!\!\perp N_t and Ai A_i iid
  • ∀n,(Z1,…,Zn)∼N(0,In) \forall n, (Z_1, \ldots ,Z_n ) \sim \mathcal{N}(0, I_n) Ai=eμz+σzZiA_i = e^{\mu_z + \sigma_z Z_i } with μz∈R \mu_z \in \mathbb{R} and σz∈R+ \sigma_z \in \mathbb{R}^+
  • (ξ( \xi , η \eta , θ)∈R3+ \theta )\in {\mathbb{R}^3}^+
  • μa=eμz+σz22 \mu_a = e^{\mu_z + \frac{\sigma_z^2}{2}}

European call price

We expose in this section three different methods to compute a call price in the Bates model:

  • Monte-Carlo
  • PIDE
  • Fourier transform (available for all models with a known characteristic function for the underlying)

For Monte-Carlo and PIDE methods, we refer the reader to the book of R. Cont and P. Tankov: “Financial Modelling with Jump Processes”.

We develop the Fourier transform in the following section.

Fourier transform method

We first recall the general Fourier transform pricing then apply it to the Bates model case.

The reader can expand/collapse for more details on Call/Put valuation using Fourier transform method

Let introduce some notations:

  • Call(t,S,K,T) Call(t, S, K, T) the price of an european call in the Bates model with maturity TT and strike KK
  • ∀t∈R+∗,St=C(t)eXt \forall t \in \mathbb{R}^{+*}, S_t = C(t) e^{X_t} the underlying price with ∀t∈R∗ \forall t \in \mathbb{R}^*, Xt X_t a random process with known moment-generating function noted χ \chi

The first step is to rewrite the european call price as follow:

Call(0,S,K,t)=1C(t)E[(St−K)+∣F0]=S0E[(eXt−KS0C(t))+∣F0]=S0E[(eXt−ek)+∣F0] \begin{equation*} Call(0, S, K, t) = \frac{1}{C(t)} \mathbb{E} [ (S_t - K)_{+} | F_{0}] = S_0 \mathbb{E} [ (e^{X_t} - \frac{K}{S_0 C(t)} )_{+} | F_{0}] = S_0 \mathbb{E} [ (e^{X_t} - e{^k})_{+} | F_{0}] \end{equation*}

with k=log(KS0C(t)) k = log ( \frac{K}{S_0 C(t)} ) .

Remark that E[(eXt−ek)+∣F0]=E[(eXt−ek(eXt−k∧1))∣F0] \mathbb{E} [ (e^{X_t} - e{^k})_{+} | F_{0}] = \mathbb{E} [ (e^{X_t} - e^{k} ( e^{X_t - k} \land 1) ) | F_{0}]

Let’s develope the conditional expectation:

E[(eXt−ek(eXt−k∧1))∣F0]=∫−∞∞exf(x)dx−e(−α+1)k∫−∞∞(1∧e−(k−x))eα(k−x)eαxf(x)dx \mathbb{E} [ (e^{X_t} - e^{k} \left( e^{X_t - k} \land 1) \right) | F_{0}] = \displaystyle \int\limits_{-\infty}^{\infty} e^{x} f(x) dx - e^{(-\alpha + 1) k} \displaystyle \int\limits_{-\infty}^{\infty} ( 1 \land e^{- (k - x) } ) e^{ \alpha (k - x)} e^{ \alpha x} f(x) dx

with f(x) f(x) the density function of XtX_t and α∈[0,1] \alpha \in \mathopen{[}0,1\mathclose{]}

We have the following result:

E[(eXt−ek)+∣F0]=χ(1)−e(−α+1)k∫−∞∞(1∧e−(k−x))eα(k−x)eαxf(x)dx \mathbb{E} [ (e^{X_t} - e{^k})_{+} | F_{0}] = \chi (1) - e^{(-\alpha + 1)k} \displaystyle \int\limits_{-\infty}^{\infty} ( 1 \land e^{- (k - x) } ) e^{ \alpha (k - x)} e^{ \alpha x} f(x) dx

Let’s note: g(x)=(1∧e−x)eαx g(x) = ( 1 \land e^{- x} ) e^{ \alpha x} and h(x)=eαxf(x) h(x) = e^{ \alpha x} f(x)

Using the property: g∗h=TF−1(TF(g∗h))=TF−1(TF(g)TF(h)) g * h = TF^{-1} ( TF (g * h) ) = TF^{-1} ( TF (g) TF (h) )

We can compute the two followings Fourier transform:

TF(g)(u)=∫−∞∞eiux(1∧e−x)eαx=∫−∞0eiux+αxdx+∫0∞eiux+x(α−1)dx=1iu+α−1iu+α−1=−1(iu+α)(iu+α−1)TF(h)(u)=∫−∞∞eiuxeαxf(x)dx=χ(α+iu) \begin{align*} TF (g) (u) & = \int\limits_{-\infty}^{\infty} e^{i u x} ( 1 \land e^{- x} ) e^{ \alpha x} \\ & = \int\limits_{-\infty}^{0} e^{i u x + \alpha x} dx + \int\limits_{0}^{\infty} e^{i u x + x (\alpha-1) } dx \\ & = \displaystyle \frac{1}{i u + \alpha} - \frac{1}{i u + \alpha -1} = \frac{-1}{(i u + \alpha) (i u + \alpha -1)} \\ TF (h) (u) & = \int\limits_{-\infty}^{\infty} e^{i u x} e^{ \alpha x} f(x) dx = \chi (\alpha + i u) \end{align*}

Therefore, we have: g∗h=−12π∫−∞∞e−kiu(iu+α)(iu+α−1)χ(α+iu)du g * h = \frac{- 1}{2 \pi} \displaystyle \int\limits_{-\infty}^{\infty} \frac{e^{-k i u }}{(i u + \alpha) (i u + \alpha -1)} \chi (\alpha + i u) du
The call price is finally:

Call(0,S,K,t)=S0χ(1)+S0ek(−α+1)2π∫−∞∞e−kiu(iu+α)(iu+α−1)χ(α+iu)du Call(0, S, K, t) = S_0 \chi(1) + S_0 \displaystyle \frac{ e^{k (-\alpha + 1)}}{2 \pi} \displaystyle \int\limits_{-\infty}^{\infty} \frac{e^{-k i u }}{(i u + \alpha) (i u + \alpha -1)} \chi (\alpha + i u) du

Because the function are Hermitian χ(α+iu)‾=χ(α−iu) \overline{\chi (\alpha + iu)} = \chi (\alpha - iu) , we have the property for a function HH: ∀x∈C\forall x \in \mathbb{C}, H(x)‾=H(−x) \overline{H(x)} = H(-x)

Let’s rewrite the integral term:

∫−∞+∞H(x)dx=∫−∞0H(x)dx+∫0+∞H(x)dx=∫0+∞H(x)dx‾+∫0+∞H(x)dx=2Re(∫0+∞H(x)dx) \displaystyle \int\limits_{-\infty}^{+\infty} H (x) dx = \int\limits_{-\infty}^{0} H (x) dx + \int\limits_{0}^{+\infty} H (x) dx = \displaystyle \overline{\int\limits_{0}^{+\infty} H (x) dx} + \int\limits_{0}^{+\infty} H (x) dx = 2 Re ( \int\limits_{0}^{+\infty} H (x) dx )

We finally have: ∫−∞∞H(x)dx=2Re∫0∞H(x)dx \displaystyle \int\limits_{-\infty}^{\infty} H (x) dx = 2 Re \int\limits_{0}^{\infty} H (x) dx .

The price at 0 of an european call on S with maturity t and strike K is:

Call(0,S,K,t)=S0χ(1)+S0ek(−α+1)π∫0∞Re(e−kiu(iu+α)(iu+α−1)χ(α+iu))du=S0χ(1)+(KC(t))(−α+1)S0απ∫0∞Re(e−kiu(iu+α)(iu+α−1)χ(α+iu))du \begin{align*} Call(0, S, K, t) = & S_0 \chi(1) + S_0 \frac{e^{k(-\alpha + 1)}}{\pi} \displaystyle \int\limits_{0}^{\infty} Re \left( \frac{e^{-k i u}}{(i u + \alpha) (i u + \alpha -1)} \chi (\alpha + i u) \right) du \\ = & S_0 \chi(1) + \displaystyle \left({\frac{K}{C(t)}}\right)^{(-\alpha + 1)} \frac{S_0^{\alpha}}{\pi} \displaystyle \int\limits_{0}^{\infty} Re \left( \frac{e^{-k i u}}{(i u + \alpha) (i u + \alpha -1)} \chi (\alpha + i u) \right) du \end{align*}

Application in the Bates model

The reader can expand/collapse for more details on Call valuation in Bates model using Fourier transform method

To increase the integral convergence, we add and remove the call price in another model, for instance in the Merton model noted CMC_{M}:

Call(0,S,K,t)=CM(0,S,K,t)+C(t)−1E[(St−K)+∣F0]−CM(0,S,K,t) Call(0, S, K, t) = C_{M}(0, S, K, t) + C(t)^{-1} \mathbb{E} [ (S_t - K)_{+} | F_{0}] - C_{M}(0, S, K, t)

Simply applying the previously found general formula, we have for Call(0,S,K,t)Call(0, S, K, t) :

CM(0,S,K,t)+S0χBates(1)+(KC(t))(−α+1)S0απ∫0∞Re(e−kiu(iu+α)(iu+α−1)χBates(α+iu))du−S0χm(1)−(KC(t))(−α+1)S0απ∫0∞Re(e−kiu(iu+α)(iu+α−1)χm(α+iu))du=CM(0,S,K,t)+(KC(t))(−α+1)S0απ∫0∞Re[e−kiu(iu+α)(iu+α−1)(χBates(α+iu)−χm(α+iu))]du \begin{align*} & C_{M}(0, S, K, t) \\ & + S_0 \chi_{Bates}(1) + \left({\frac{K}{C(t)}}\right)^{(-\alpha + 1)} \frac{S_0^{\alpha} }{\pi} \displaystyle \int\limits_{0}^{\infty} Re \left( \frac{e^{-k i u}}{(i u + \alpha) (i u + \alpha -1)} \chi_{Bates} (\alpha + i u) \right) du \\ & - S_0 \chi_{m}(1) - \left({\frac{K}{C(t)}}\right)^{(-\alpha + 1)} \frac{S_0^{\alpha} }{\pi} \displaystyle \int\limits_{0}^{\infty} Re \left( \frac{e^{-k i u}}{(i u + \alpha) (i u + \alpha -1)} \chi_{m} (\alpha + i u) \right) du \\ = & C_{M}(0, S, K, t) \\ & + \left({\frac{K}{C(t)}}\right)^{(-\alpha + 1)} \frac{S_0^{\alpha} }{\pi} \displaystyle \int\limits_{0}^{\infty} Re \left[ \frac{e^{-k i u}}{(i u + \alpha) (i u + \alpha -1)} ( \chi_{Bates} (\alpha + i u) - \chi_{m}(\alpha + i u) ) \right] du \end{align*}

The first term is computed using the series expansion found in the section dedicated to the Merton model.

Expression of χBates(u)\chi_{Bates}(u) and χm(u)\chi_{m}(u) ∀u∈C\forall u \in \mathbb{C}

Computation of χm(u)\chi_{m}(u), ∀u∈C\forall u \in \mathbb{C}

∀u∈C,χm(u)=χmd(u)×χmc(u)∀u∈C,χmc(u)=χmc(u)=χBS(u)=e−σ2t2(u−u2)∀u∈C,χmd(u)=E[eJt]=eλt(χa(u)−1)=eλt(eu(μz+uσz22)−1) \begin{align*} \forall u \in \mathbb{C}, \chi_m(u) & = \chi_m^d(u) \times \chi_m^c(u) \\ \forall u \in \mathbb{C}, \chi_m^c(u) & = \chi_m^c(u) = \chi_{BS}(u) = e^{ - \frac{\sigma^2 t}{2}(u-u^2)} \\ \forall u \in \mathbb{C}, \chi_{m}^d(u) & = \mathbb{E} [ e^{J_t} ] = e^{ \lambda t ( \chi_a(u) - 1 )} \\ & = \displaystyle e^{ \lambda t \left( e^{u ( \mu_z + \frac{u \sigma_z^2}{2} )} - 1 \right)} \end{align*}

With χa\chi_a the Moment-Generating function of the random variable AiA_i.

We note ∀u∈C\forall u \in \mathbb{C}, χmd(u)=χJump(u) \displaystyle \chi_{m}^d(u) = \chi_{Jump}(u)

Computation of χBates(u)\chi_{Bates}(u), ∀u∈C\forall u \in \mathbb{C}

∀u∈C,χBates(u)=χBatesc(u)×χBatesd(u)∀u∈C,χBatesd(u)=χJump(u)=E[eJt]=exp(λT(euμz+u2σz22−1))∀u∈C,χBatesc(u)=χHeston(u) \begin{align*} \forall u \in \mathbb{C}, \chi_{Bates}(u) & = \chi_{Bates}^c(u) \times \chi_{Bates}^d(u) \\ \forall u \in \mathbb{C},\chi_{Bates}^d(u) & = \chi_{Jump}(u) = \mathbb{E} [ e^{J_t} ] = exp\left(\lambda T( e^{u \mu_z + u^2 \frac{\sigma_z^2 }{2} } - 1)\right) \\ \forall u \in \mathbb{C}, \chi_{Bates}^c(u) & = \chi_{Heston}(u) \end{align*}

First of all, let’s note X=log(S) X = log(S) . Applying Itô formula to log(S) log(S) , we have:

dXt=(rt+λ(1−mua)−12Vt)dt+VtdWts dX_t = (r_t + \lambda(1 - mu_a) - \frac{1}{2}V_t )dt + \sqrt{V_t} dW_t^s

Let’s note ∀t∈R+ \forall t \in \mathbb{R}^{+} Mt=f(Xtc,Vt,t) M_t = f(X_t^c, V_t, t ) with ∀(x,v,t)∈(R+)3 \forall (x,v,t) \in (\mathbb{R}^{+})^3 f(x,v,t)=E[euXtc∣Xtc=x,Vt=v] f(x, v, t ) = \mathbb{E} [ e^{u X_t^c} |X_t^c = x , V_t = v ]

For the sake of clarity, let’s introduce f=f(Xtc,Vt,t) f = f(X_t^c, V_t, t) and let’s apply the Itô formula for continuous processes to Mt M_t .

dMt=∂f∂xdXtc+∂f∂vdVt+∂f∂tdt+12[∂2f∂x2d[Xc]t+∂2f∂v2d[V]t+2∂2f∂x∂vd[Xc,V]t] d M_t = \displaystyle \frac{\partial f }{\partial x } d X_t^c + \frac{\partial f }{\partial v } d V_t + \frac{\partial f }{\partial t } d t + \frac{1}{2} \left[ \frac{\partial^2 f }{\partial x^2 } d[ X^c ]_t + \frac{\partial^2 f }{\partial v^2 } d[ V ]_t + 2 \frac{\partial^2 f }{\partial x \partial v } d[ X^c , V ]_t \right]

Let’s compute each terms:

dXtc=(rt+λ(1−μA)−Vt2)dt+VtdWtsdVt=ξ(η−Vt)dt+θVtdWtvd[Xc]t=Vtdtd[V]t=θ2Vtdtd[Xc,V]t=ρθVtdt \begin{align*} d X_t^c & = \displaystyle (r_t + \lambda (1 - \mu_A) - \frac{V_t}{2}) dt + \sqrt{V_t} d W_t^s \\ d V_t & = \xi (\eta - V_{t}) dt + \theta \sqrt{V_{t}} d W_t^v \\ d[ X^c ]_t & = V_t dt \\ d[ V ]_t & = \theta^2 V_t dt \\ d[ X^c , V ]_t & = \rho \theta V_t dt \end{align*}

Substituting, we have:

dMt=∂f∂x((rt+λ(1−μA)−Vt2)dt+VtdWts)+∂f∂v(ξ(η−Vt)dt+θVtdWtv)+∂f∂tdt+12[∂2f∂x2Vt+∂2f∂v2θ2Vt+2∂2f∂x∂vρθVt]dt=[∂f∂x(rt+λ(1−μA)−Vt2)+∂f∂vξ(η−Vt)+∂f∂t+12∂2f∂x2Vt+12∂2f∂v2θ2Vt+∂2f∂x∂vρθVt]dt+θVtdWtv+VtdWts \begin{align*} d M_t = & \displaystyle \frac{\partial f }{\partial x } ( (r_t + \lambda (1 - \mu_A) - \frac{V_t}{2}) dt + \sqrt{V_t} d W_t^s ) + \frac{\partial f }{\partial v } ( \xi (\eta - V_{t}) dt + \theta \sqrt{V_{t}} d W_t^v ) + \frac{\partial f }{\partial t } d t \\ & \displaystyle + \frac{1}{2} [ \frac{\partial^2 f }{\partial x^2 } V_t + \frac{\partial^2 f }{\partial v^2 } \theta^2 V_t + 2 \frac{\partial^2 f }{\partial x \partial v } \rho \theta V_t ] dt \\ = & [ \displaystyle \frac{\partial f }{\partial x } (r_t + \lambda (1 - \mu_A) - \frac{V_t}{2}) + \frac{\partial f }{\partial v } \xi (\eta - V_{t}) + \frac{\partial f }{\partial t } \\ & + \frac{1}{2}\frac{\partial^2 f }{\partial x^2 } V_t + \frac{1}{2}\frac{\partial^2 f }{\partial v^2 } \theta^2 V_t + \frac{\partial^2 f }{\partial x \partial v } \rho \theta V_t ] dt + \displaystyle \theta \sqrt{V_{t}} d W_t^v + \sqrt{V_{t}} d W_t^s \end{align*}

Moreover, we know that f(X⋅c,V⋅,⋅)f(X_{\cdot}^c, V_{\cdot}, \cdot) has to be a martingale. The drift must be equals to zero because f(X⋅c,V⋅,⋅)f(X_{\cdot}^c, V_{\cdot}, \cdot) is continue.

∂f∂x(rt+λ(1−μA)−Vt2)+∂f∂vξ(η−Vt)+∂f∂t+12∂2f∂x2Vt+12∂2f∂v2θ2Vt+∂2f∂x∂vρθVt=0 \displaystyle \frac{\partial f }{\partial x } (r_t + \lambda (1 - \mu_A) - \frac{V_t}{2}) + \frac{\partial f }{\partial v } \xi (\eta - V_{t}) + \frac{\partial f }{\partial t } + \frac{1}{2}\frac{\partial^2 f }{\partial x^2 } V_t + \frac{1}{2}\frac{\partial^2 f }{\partial v^2 } \theta^2 V_t + \frac{\partial^2 f }{\partial x \partial v} \rho \theta V_t = 0

with the terminal condition f(x,v,t)=eux f(x,v,t) = e^{ux}

We suppose the solution can be written as follow: f(x,v,t)=eA(T−t)+vB(T−t)+ux f(x, v, t) = e^{ A(T-t) + v B(T-t) + u x } with initial condition A(0)=0 A(0) = 0 and B(0)=0 B(0) = 0 . We have:

∂f∂x=uf∂f∂v=B(T−t)f∂f∂t=−∂A∂t(T−t)+−Vt∂B∂t(T−t)∂2f∂x2=u2f∂2f∂v2=B(T−t)2f∂2f∂x∂v=−uB(T−t)f \begin{align*} \displaystyle \frac{\partial f }{\partial x } & = u f \\ \frac{\partial f }{\partial v } & = B(T-t) f \\ \frac{\partial f }{\partial t } & = - \frac{\partial A }{\partial t }(T-t) + -V_t \frac{\partial B }{\partial t }(T-t) \\ \frac{\partial^2 f }{\partial x^2 } & = u^2 f \\ \displaystyle \frac{\partial^2 f }{\partial v^2 } & = B(T-t)^2 f \\ \frac{\partial^2 f }{\partial x \partial v } & = - u B(T-t) f \end{align*}

The equation can be written by substituting:

uf(rt+λ(1−μA)−Vt2)+B(t)fξ(η−Vt)−(∂A∂t(t)+Vt∂B∂t(t))+12u2fVt+12B(t)2fθ2Vt−uB(t)fρθVt=0 \begin{align*} & \displaystyle u f (r_t + \lambda (1 - \mu_A) - \frac{V_t}{2}) + B(t) f \xi (\eta - V_{t}) - (\frac{\partial A }{\partial t }(t) + V_t \frac{\partial B }{\partial t }(t) ) \\ & + \frac{1}{2}u^2 f V_t + \frac{1}{2}B(t)^2 f \theta^2 V_t - u B(t) f \rho \theta V_t = 0 \end{align*} u(rt+λ(1−μA))+B(t)ξη−∂A∂t(t)+[−∂B∂t(t)−B(t)ξ+12u2+12B(t)2θ2−uB(t)ρθ−u2]Vt=0 \begin{align*} & \displaystyle u(r_t + \lambda (1 - \mu_A)) + B(t)\xi \eta - \frac{\partial A }{\partial t }(t) \\ & + \left[ - \frac{\partial B }{\partial t }(t) - B(t)\xi + \frac{1}{2}u^2 + \frac{1}{2}B(t)^2 \theta^2 - u B(t) \rho \theta - \frac{u}{2} \right] V_t = 0 \end{align*}

The two solutions of our problem are:

∂A∂t(t)=B(t)ξη+u(rt+λ(1−μA))∂B∂t(t)=12θ2B(t)2−ξB(t)−ρθuB(t)+12u2−12u=12θ2B(t)2−(ρθu+ξ)B(t)−u2(1−u) \begin{align*} \displaystyle \frac{\partial A }{\partial t }(t) & = B(t) \xi \eta + u (r_t + \lambda (1 - \mu_A)) \\ \displaystyle \frac{\partial B }{\partial t }(t) & = \frac{1}{2}\theta^2 B(t)^2 - \xi B(t) - \rho \theta u B(t) + \frac{1}{2}u^2 - \frac{1}{2}u = \frac{1}{2}\theta^2 B(t)^2 - ( \rho \theta u + \xi ) B(t) - \frac{u}{2}( 1 - u ) \end{align*}

The first one is an integral computation. The second one is a Riccati equation.

The solutions of our two equations can be written as follow:

A(t)=u(1log(C(t))+tλ(1−μA))+ξηt(ξ+ρθu)θ2−2ξηθ2log(cosh(γt2)−ξ+ρθuγsin(γt2))B(t)=−u−u2γcoth(γt2)−ξ−ρθu \begin{align*} \displaystyle A(t) & = u \left( \frac{1}{log ( C(t))} + t \lambda ( 1 - \mu_A) \right) + \frac{\xi \eta t (\xi + \rho \theta u)}{\theta^2} - \frac{2 \xi \eta}{\theta^2} log \left( cosh ( \frac{\gamma t}{2} ) - \frac{\xi + \rho \theta u}{\gamma} sin ( \frac{\gamma t}{2} ) \right) \\ \displaystyle B(t) & = - \frac{u - u^2}{\gamma coth ( \frac{\gamma t}{2} ) - \xi - \rho \theta u } \end{align*}

with γ=θ2(u−u2)+(ξ+ρθu)2 \gamma = \displaystyle \sqrt{\theta^2 (u - u^2) + (\xi + \rho \theta u)^2}

We have the log characteristic function:

log(χBatesc(u))=u(1log(C(t))+tλ(1−μA))+ξηt(ξ+ρθu)θ2−2ξηθ2log(cosh(γt2)−ξ+ρθuγsin(γt2))−V0u−u2γcoth(γt2)−ξ−ρθu+uX0 \begin{align*} log(\chi_{Bates}^c(u)) &= u \left( \frac{1}{log ( C(t) )} + t \lambda ( 1 - \mu_A)\right) + \frac{\xi \eta t (\xi + \rho \theta u)}{\theta^2}\\ &- \frac{2 \xi \eta}{\theta^2} log \left(cosh ( \frac{\gamma t}{2}) - \frac{\xi + \rho \theta u}{\gamma} sin ( \frac{\gamma t}{2} ) \right) - V_0 \frac{u - u^2}{\gamma coth ( \frac{\gamma t}{2} ) - \xi - \rho \theta u } + u X_0 \end{align*}

and the call price in the Bates model is given by:

Call(0,S,K,t)=CM(0,S,K,t)−KπC(t)∫0∞Re[e−k(α+iu)χJ(α+iu)(iu+α)(iu+α−1)(χHeston(α+iu)−χBS(α+iu))]du \begin{align*} Call(0, S, K, t) &= C_M(0, S, K, t)\\ & - \frac{K}{\pi C(t)} \displaystyle \int\limits_{0}^{\infty} Re \left[ \frac{e^{-k( \alpha + i u )} \chi_J (\alpha + i u) }{(i u + \alpha) (i u + \alpha -1)} \left( \chi_{Heston}(\alpha + i u) - \chi_{BS}(\alpha + i u) \right) \right] du \end{align*}

Control variate choice

To compute the Bates’ call price, we can choose the Merton, Black-Scholes, or no control variate to increase the accuracy.

Remark: Using a Black-Scholes or Merton call price as control variate, we have to choose the volatility σ \sigma parameter.

  • Let’s identify σ\sigma by choosing θ=0\theta = 0. We have ∀t∈R+∗ \forall t \in \mathbb{R}^{+*} : d(eξtVt)=Vtξeξtdt+eξtdVt=Vtξeξtdt+eξtξ(η−Vt)dt=eξtξηdteξtVt−V0=∫0teξtξηdt  ⟺  Vt=e−ξt(V0−η)+η \begin{align*} \displaystyle d( e^{\xi t} V_t ) & = V_t \xi e^{\xi t} dt + e^{\xi t} dV_t = V_t \xi e^{\xi t} dt + e^{\xi t} \xi (\eta - V_{t}) dt =e^{\xi t} \xi \eta dt \\ \displaystyle e^{\xi t} V_t - V_0 & = \int\limits_0^t e^{\xi t} \xi \eta dt \iff V_t = e^{-\xi t} (V_0 - \eta) + \eta \end{align*}
  • σ \sigma can be chosen as the squared root of the terminal expected variance: σ=e−ξt(V0−η)+η \sigma = \displaystyle \sqrt{e^{-\xi t} (V_0 - \eta) + \eta}
  • σ\sigma can be chosen as the squared root of the integrated variance without noise: σ2=∫0tVsds=1t[(e−ξt−1)(η−V0)ξ+ηt]  ⟺  σ=(e−ξt−1)(η−V0)ξt+η \sigma^2 = \displaystyle \int\limits_0^t V_s ds = \frac{1}{t} [ \frac{(e^{-\xi t} - 1)(\eta - V_0)}{\xi} +\eta t ] \iff \sigma = \sqrt{\frac{(e^{-\xi t} - 1)(\eta - V_0)}{\xi t} +\eta}

Calibration methods

We develop two calibration methods for the Bates model:

  • In one step and using weight functions
  • In two steps: Firstly a set of parameters is found using Merton and Heston calibration and secondly a local optimisation is performed

One step method

We minimise the imply/model volatilities in L2L^2 norm, as explain in the Merton calibration section. Let’s define a weight function applied respectively to strikes and to maturities.

  • Weight function for strikes:

    ∀x∈R+∗, \forall x \in \mathbb{R}^{+*}, f(x)=e−((x−μ)σ)2σ2π f(x) = \displaystyle \frac{ e^{ - \left( \frac{(x - \mu)}{\sigma} \right)^2 }}{\sigma \sqrt{2 \pi}} with μ=S0C(t)\mu = S_0 C(t) ,
    σ=scale×12(kmin2+kmax2)−14(kmin+kmax)2 \sigma = scale \times \sqrt{ \frac{1}{2}(k_{min}^2 + k_{max}^2) - \frac{1}{4} (k_{min} + k_{max} )^2 }

    kmin k_{min} minimum strike and kmax k_{max} maximum strike for a fixed T∈R+∗ T \in \mathbb{R}^{+*} and scale∈R+∗ scale \in \mathbb{R}^{+*} a scale factor

  • Weight function for maturities:

    ∀x∈R+∗, \forall x \in \mathbb{R}^{+*}, f(x)=(xx0)px0ep(x−x0) f(x) = \displaystyle (\frac{x}{x_0})^{p x_0} e^{ p(x-x_0) } with x0 x_0 the centered maturity and pp the bandwidth

Two steps method

We inspect a first solution in the set of Bates parameters.

For the height parameters, we have:

  • For λ\lambda, μz\mu_z, σz\sigma_z and σ\sigma:

    We calibrate those parameters in the Merton model for maturities lower than one year. Remark that σ\sigma is the volatility in the Merton model and not a Bates parameter

  • For V0V_0, ξ\xi, ν\nu, θ\theta and ρ\rho:

    We calibrate those parameters in the Heston model for maturities greater than one year. Remark that v0v_0 can be taken equals to σ2 \sigma^2

The second step use the previously found parameters as initial solution for a local minimisation.

Parameters impact

To expose the parameters influence, we expose the implied volatility surface in function of the model parameters. Don’t hesitate to expand this section

Let’s have the following parameters:

  • s0 = 100
  • r = 0.2
  • lambda = 0.5
  • v0 = 0.4
  • mu_z = -0.2
  • sigma_z = 0.2
  • kappa = 0.5
  • theta = 0.4
  • xi = 0.9
  • rho = -0.7
This will display an animated GIF

Conclusion

We presented in this post some theoretical aspects on discontinuous processes. We also exposed a generic formulation for jump diffusion model using Lévy process. We investigated two specific models: Merton and Bates models. The dynamic, call pricing, calibration and influence of parameters are exposed for each models. Numerical experiments are provided.

This methodology has been chosen to be implemented in LexiFi Apropos.