MatheLinuxMatheLinux
Curriculum
Sign inStart free
BSDE, XVA & Credit Risk
0%

0 of 14 lessons complete

Lessons

  • Backward Stochastic Differential Equations8m
  • Nonlinear Feynman–Kac8m
  • The Comparison Theorem & Properties8m
  • BSDEs in Pricing & Hedging8m
  • Modeling Default: Hazard Rates8m
  • Credit Exposure: EE, EPE, PFE8m
  • Wrong-Way Risk8m
  • CVA & DVA8m
  • FVA: Funding Valuation Adjustment8m
  • KVA & MVA8m
  • The Unified XVA Equation8m
  • Monte Carlo Exposure Simulation8m
  • Regression & Longstaff–Schwartz for Conditional Expectations8m
  • Deep BSDE Solvers8m

MatheLinux — quantitative finance, taught rigorously.

Course content credited to Ibrahim Lanre Adedimeji.

← BSDE, XVA & Credit Risk

A backward stochastic differential equation (BSDE) runs an ordinary stochastic differential equation in reverse. Instead of fixing where the process starts and letting it diffuse forward, we fix where it must end and ask what dynamics carry it there. This is exactly the situation a derivatives desk faces: we know the payoff at maturity, and we want today's value plus the trading strategy that delivers that payoff. The terminal condition is data; the present is the unknown.

Intuition

Forward stochastic dynamics answer the question "given today, what happens next?" Pricing asks the opposite: "given a contractual payoff at time TTT, what is the fair value now, and how must I trade in between?" A naive attempt to integrate a stochastic differential equation backward fails because the answer would depend on the future of the Brownian motion, violating the requirement that a price be known today. BSDE theory resolves this by introducing a second unknown, the control process ZZZ, whose job is to absorb the unpredictable part so that the solution stays adapted, that is, depends only on information available up to the current time.

Formal definition

Let WWW be a ddd-dimensional Brownian motion generating the filtration, let the terminal time be TT, let the terminal value be square-integrable and known at time , and let the driver map time, value, and control to a real number. A solution to the BSDE is a pair of adapted processes satisfying

Knowledge check

Q1. In the BSDE solution pair, what is the financial meaning of the control process Z?

Q2. Which condition on the driver f is the standard hypothesis guaranteeing a unique square-integrable adapted solution?

Q3. Why must a second unknown Z be introduced rather than just solving for Y?

Q4. For the linear driver f(t,y,z) = -r y, what is the resulting value process Y_t?

Problems

P1

A BSDE has driver f(t,y,z) = -r y with constant continuously-compounded rate r = 0.05, maturity T = 2 years, and a deterministic terminal payoff xi = 100. Since the payoff is deterministic, compute the present value Y_0. Give the decimal value.

P2

Show symbolically that for the linear BSDE with driver f(t,y,z) = -r y, the candidate Y_t = E[exp(-r(T-t)) xi | F_t] satisfies the BSDE, and identify Z_t.

Mark this lesson complete to track progress
T
ξ\xiξ
TTT
fff
(Y,Z)(Y,Z)(Y,Z)

−dYt=f(t,Yt,Zt) dt−Zt dWt,YT=ξ.-dY_t = f(t,Y_t,Z_t)\,dt - Z_t\,dW_t, \qquad Y_T = \xi.−dYt​=f(t,Yt​,Zt​)dt−Zt​dWt​,YT​=ξ.

Written in integral form, which is the mathematically precise statement, this reads

Yt=ξ+∫tTf(s,Ys,Zs) ds−∫tTZs dWs.Y_t = \xi + \int_t^T f(s,Y_s,Z_s)\,ds - \int_t^T Z_s\,dW_s.Yt​=ξ+∫tT​f(s,Ys​,Zs​)ds−∫tT​Zs​dWs​.

Here YYY is the value process and ZZZ is the control process. The solution is the pair that makes both lines hold while keeping YYY and ZZZ adapted.

The role of Z

The stochastic integral term has zero conditional expectation, so taking conditional expectations of the integral form gives

Yt=E ⁣[ξ+∫tTf(s,Ys,Zs) ds ∣ Ft].Y_t = \mathbb{E}\!\left[\xi + \int_t^T f(s,Y_s,Z_s)\,ds \,\Big|\, \mathcal{F}_t\right].Yt​=E[ξ+∫tT​f(s,Ys​,Zs​)ds​Ft​].

This would seem to determine YYY on its own, but the right side still contains ZZZ through the driver. The role of ZZZ is structural: it is the martingale representation of the value process. The martingale MtM_tMt​ equal to the conditional expectation of ξ\xiξ plus the running driver integral can be written as a stochastic integral against WWW, and its integrand is precisely ZZZ. In hedging language, ZZZ is the sensitivity of value to the driving noise, the quantity you trade to neutralize risk. Without ZZZ there is no way to make the terminal condition hold while staying adapted; with it, the equation closes.

Existence and uniqueness under Lipschitz f

The foundational result of Pardoux and Peng states that if ξ\xiξ is square-integrable, the driver f(⋅,0,0)f(\cdot,0,0)f(⋅,0,0) is square-integrable in time, and fff is uniformly Lipschitz in (y,z)(y,z)(y,z), meaning there is a constant KKK with

∣f(t,y1,z1)−f(t,y2,z2)∣≤K(∣y1−y2∣+∣z1−z2∣),|f(t,y_1,z_1)-f(t,y_2,z_2)| \le K\big(|y_1-y_2| + |z_1-z_2|\big),∣f(t,y1​,z1​)−f(t,y2​,z2​)∣≤K(∣y1​−y2​∣+∣z1​−z2​∣),

then the BSDE has a unique adapted square-integrable solution pair (Y,Z)(Y,Z)(Y,Z).

Derivation sketch

The proof is a fixed-point argument. Define a map that takes a candidate pair, freezes it inside the driver, and solves the resulting linear problem by martingale representation to produce a new pair. One shows this map is a contraction in a weighted norm of the form

∥(Y,Z)∥β2=E ⁣[∫0Teβt(∣Yt∣2+∣Zt∣2) dt],\|(Y,Z)\|_\beta^2 = \mathbb{E}\!\left[\int_0^T e^{\beta t}\big(|Y_t|^2 + |Z_t|^2\big)\,dt\right],∥(Y,Z)∥β2​=E[∫0T​eβt(∣Yt​∣2+∣Zt​∣2)dt],

where choosing the exponential weight β\betaβ large enough relative to KKK makes the Lipschitz contribution shrink below one. By the Banach fixed-point theorem the map has a unique fixed point, which is the unique solution. The exponential weighting is the key trick: it tames the linear growth coming from the Lipschitz bound.

Worked example

Take the linear driver f(t,y,z)=−ryf(t,y,z) = -r yf(t,y,z)=−ry with constant rate rrr and terminal value ξ\xiξ. The equation is −dYt=−rYt dt−Zt dWt-dY_t = -rY_t\,dt - Z_t\,dW_t−dYt​=−rYt​dt−Zt​dWt​. Guess Yt=E[e−r(T−t)ξ∣Ft]Y_t = \mathbb{E}[e^{-r(T-t)}\xi \mid \mathcal{F}_t]Yt​=E[e−r(T−t)ξ∣Ft. Differentiating, the deterministic drift contributes rYt dtrY_t\,dtrYt​dt, matching the driver, and the martingale part supplies ZZZ through its representation. So the value process is the discounted conditional expectation of the payoff, the textbook risk-neutral price, and ZZZ is its diffusion coefficient, the delta. This recovers classical pricing as the linear special case of a BSDE.

Why quants care

BSDEs are the unifying language of modern valuation. The linear case reproduces risk-neutral pricing, but the real power appears when the driver is nonlinear: differential funding rates, collateral, credit charges, and trading constraints all enter through fff, and the same existence theory still guarantees a well-posed price and hedge. The control ZZZ is not an abstraction; it is the hedge ratio the desk actually trades. Understanding the (Y,Z)(Y,Z)(Y,Z) pair is the conceptual entry point to the entire valuation-adjustment framework that follows.

Adapted from Pardoux–Peng and Crépey — Counterparty Risk and Funding.

​
]