Stochastic Partial Differential Equations : Why Classical PDEs Fail in Markets
Ibrahim Lanre Adedimeji13 min read·Just now--
For decades, financial markets have been modeled with the elegance of classical partial differential equations. From the Black–Scholes equation to diffusion-based pricing frameworks, the underlying assumption has been deceptively simple: uncertainty can be averaged out, and randomness behaves smoothly over time.
But real markets don’t move smoothly.
Prices jump , Volatility clusters, Liquidity vanishes without warning. Entire regimes shift in ways no deterministic surface can anticipate. The neat, continuous world described by classical PDEs begins to fracture the moment it meets high-frequency data or stressed market conditions. What was once a clean diffusion becomes a noisy, path-dependent, and often discontinuous reality.
This mismatch is not just a technical inconvenience :it is a structural failure of the modeling paradigm. Classical PDEs are built on assumptions of continuity, Markovian dynamics, and well-behaved noise. Markets, on the other hand, are driven by memory, feedback loops, and randomness that evolves with time itself.
This is where stochastic partial differential equations (SPDEs) enter the picture.
Rather than treating randomness as an external input, SPDEs embed it directly into the evolution of the system. The solution itself becomes random, capable of capturing not just the average behavior of prices, but the evolving uncertainty around them. In doing so, SPDEs provide a framework that is fundamentally closer to how markets actually behave irregular, adaptive, and inherently stochastic at every scale.
In this article, we will explore why classical PDEs fall short in financial modeling, and how SPDEs offer a more faithful mathematical lens for understanding modern markets.
1)The classical promise: Black-Scholes and the illusion of certainty
BLACK–SCHOLES AS A HEAT EQUATION IN DISGUISE
The Black–Scholes PDE:
∂V/∂t + (1/2)σ²S² · ∂²V/∂S² + rS · ∂V/∂S − rV = 0
CHANGE OF VARIABLES
Introduce three substitutions:
x = ln(S) log-price (space variable)
τ = T − t time-to-expiry (time variable)
V = e^(αx + βτ) · u rescaled price
The third substitution absorbs the drift and discounting
terms. α and β are constants to be determined.
Chain rule for the partial derivatives:
∂V/∂t = −∂V/∂τ
= −e^(αx+βτ) · (βu + ∂u/∂τ)
∂V/∂S = (1/S) · ∂V/∂x
= (1/S) · e^(αx+βτ) · (αu + ∂u/∂x)
∂²V/∂S² = (1/S²) · (∂²V/∂x² − ∂V/∂x)
= (1/S²) · e^(αx+βτ) · (α²u + 2α·∂u/∂x + ∂²u/∂x²
− αu − ∂u/∂x)
= (1/S²) · e^(αx+βτ) · ((α²−α)u + (2α−1)∂u/∂x
+ ∂²u/∂x²)
SUBSTITUTE INTO THE PDE
Substitute into BS, note S² · (1/S²) = 1 and S · (1/S) = 1,
then divide every term by e^(αx+βτ) ≠ 0:
−(βu + ∂u/∂τ)
+ (1/2)σ² · ((α²−α)u + (2α−1)∂u/∂x + ∂²u/∂x²)
+ r · (αu + ∂u/∂x)
− r · u
= 0
Rearrange by grouping ∂u/∂τ on the left:
∂u/∂τ = (1/2)σ² · ∂²u/∂x²
+ [ (1/2)σ²(2α−1) + r ] · ∂u/∂x
+ [ −β + (1/2)σ²(α²−α) + rα − r ] · u
For this to reduce to a pure heat equation, the coefficients
of ∂u/∂x and u must both vanish.
SOLVE FOR α
Set the coefficient of ∂u/∂x equal to zero:
(1/2)σ²(2α − 1) + r = 0
σ²α − σ²/2 + r = 0
σ²α = σ²/2 − r
α = 1/2 − r/σ²
STEP 4 — SOLVE FOR β
Set the coefficient of u equal to zero:
−β + (1/2)σ²(α² − α) + rα − r = 0
β = (1/2)σ²(α² − α) + rα − r
Substitute α = 1/2 − r/σ²:
α² − α = (1/2 − r/σ²)² − (1/2 − r/σ²)
= 1/4 − r/σ² + r²/σ⁴ − 1/2 + r/σ²
= r²/σ⁴ − 1/4
(1/2)σ²(α² − α) = (1/2)σ²(r²/σ⁴ − 1/4)
= r²/(2σ²) − σ²/8
rα = r(1/2 − r/σ²)
= r/2 − r²/σ²
Collecting all three terms:
β = r²/(2σ²) − σ²/8 + r/2 − r²/σ² − r
= −r²/(2σ²) − r/2 − σ²/8
This factors cleanly as:
β = −(2r + σ²)² / (8σ²)
THE RESULT
With both coefficients zero, the PDE reduces exactly to:
∂u/∂τ = (1/2)σ² · ∂²u/∂x²
This is the classical heat equation, where:
u(x, τ) rescaled option value
x log-price : plays the role of position
τ time-to-expiry : plays the role of time
(1/2)σ² diffusion coefficient :thermal diffusivity
The fundamental solution is a Gaussian kernel.
That is precisely why the final formula involves N(d₁) and N(d₂):
the normal CDF is the integrated heat kernel, translated back
into the original (S, t) coordinates.
d₁ = [ ln(S/K) + (r + σ²/2)(T−t) ] / σ√(T−t)
d₂ = d₁ − σ√(T−t)
Call: C = S·N(d₁) − K·e^(−r(T−t))·N(d₂)
Put: P = K·e^(−r(T−t))·N(−d₂) − S·N(−d₁)
THE THREE ASSUMPTIONS HOLDING THIS TOGETHER
1. CONSTANT VOLATILITY
σ is a fixed real number : not a function of S or t.
This keeps the diffusion coefficient (1/2)σ² constant,
making the equation a true heat equation.
If σ = σ(S, t), the coefficient varies and the
closed-form solution no longer exists.
2. CONTINUOUS FRICTIONLESS TRADING
The hedge ratio Δ = ∂V/∂S can be adjusted continuously
at zero cost. This is what allows all randomness (dS terms)
to cancel in the replicating portfolio, leaving a
deterministic, riskless equation.
Without it, residual risk remains and the PDE breaks down.
3. CONTINUOUS PATHS :NO JUMPS
The asset S follows a Wiener process: paths are continuous
almost surely, and (dW_t)² = dt holds exactly.
This identity is what produces the second-order term
(1/2)σ²S²·∂²V/∂S² via Itô's lemma.
If S can jump, that term gains a non-local integral
and the heat equation structure is destroyed.
2) What PDEs can’t see : Three specific market realities that classical PDEs :
- Stochastic volatility : is not a number you write down once. It clusters, spikes, and mean-reverts. Plot implied vol against any strike and you get a smile, never a flat line. The PDE has one diffusion coefficient. Markets have a whole surface.
- Jump discontinuities : when a stock gaps 15% on earnings, prices didn’t travel there, they teleported. Itô’s lemma needs continuity to work. A jump breaks the hedge instantly, by a finite amount, with zero warning. The PDE has no term for that.
- Regime changes : the correlation structure of 2019 and March 2020 are not the same object. Fixed coefficients assume the medium doesn’t change. But r, σ, and correlations are emergent properties of whatever regime you’re in and regimes break without asking permission.
3) The Realm of SPDE’s :
ENTER THE SPDE : RANDOMNESS ALL THE WAY DOWN
In a classical PDE, the equation is deterministic.
You solve it, you get a surface V(S, t) one price
for every (S, t) pair. Uncertainty lives outside
the equation, in the inputs.
In an SPDE, noise is structural.
It lives inside the equation itself, multiplying
the solution, the state, and time simultaneously.
You do not solve for a surface.
You solve for a random field ,an infinite family
of surfaces, one for each realisation of the noise.
THE GENERAL FORM
dV(x, t) = L[V](x, t) dt + σ(V, x, t) dW(x, t)
WHAT EACH PIECE MEANS :
V(x, t)
─────────────────────────────────────────────────────
The unknown. Not a single number but a random field :
a function of both position x and time t that is
itself stochastic. Every realisation of the driving
noise W gives a different V(x, t).
L[V](x, t)
─────────────────────────────────────────────────────
A differential operator acting on V.
This is the deterministic skeleton of the equation —
the part inherited from classical PDEs.
In financial terms it carries drift, discounting,
and diffusion in the underlying.
Examples of L:
L[V] = (1/2)σ²S²∂²V/∂S² + rS∂V/∂S − rV
(Black–Scholes operator)
L[V] = (1/2)v·S²∂²V/∂S² + κ(θ−v)∂V/∂v + ...
(Heston operator, v = stochastic variance)
σ(V, x, t)
─────────────────────────────────────────────────────
The diffusion coefficient of the noise term.
This is what makes the equation stochastic
and what makes it an SPDE rather than a standard SDE.
Crucially, σ depends on:
V — the solution itself (nonlinear noise)
x — position in state space (spatially varying)
t — time (time-varying)
When σ depends on V, the noise amplitude changes
with the solution the equation is said to have
multiplicative noise. This is the generic case
in financial models.
When σ is constant, the noise is additive
a special, simpler case.
dW(x, t)
─────────────────────────────────────────────────────
A space-time white noise process the SPDE
generalisation of the Wiener process dW_t.
Formally: dW(x, t) = W(dx, dt)
where W(A, t) is a Brownian sheet:
· For fixed x, t ↦ W(x, t) is a Brownian motion
· For fixed t, x ↦ W(x, t) is spatially random
· Increments are independent across disjoint
regions of (x, t) space
Covariance structure:
E[ dW(x,t) · dW(y,s) ] = Q(x, y) · δ(t−s) dt
where Q(x, y) is the spatial covariance kernel.
When Q(x, y) = δ(x−y): pure white noise in space
When Q(x, y) = e^(−|x−y|/ℓ): spatially correlated
noise with length
scale ℓ
THE KEY CONCEPTUAL SHIFT
Classical PDE │ SPDE
───────────────────────┼──────────────────────────────
One equation │ One equation
One solution surface │ A distribution of surfaces
V(S, t) is a function │ V(x, t) is a random field
σ enters as a │ σ(V, x, t) is inside
parameter │ the dynamics
Noise is external │ Noise is structural
Solve once │ Every ω ∈ Ω gives a
│ different realisation
In other words:
A classical PDE asks :
given these inputs, what is the price?
An SPDE asks :
given these inputs and this realisation
of the noise, what is the price?
And what is the full distribution
across all realisations?
SOLUTION THEORY : WHAT IT MEANS TO SOLVE AN SPDE
A solution V to the SPDE is a stochastic process
adapted to the filtration F_t generated by W, such
that for all t:
V(x,t) = V(x,0)
+ ∫₀ᵗ L[V](x,s) ds
+ ∫₀ᵗ σ(V,x,s) dW(x,s)
The second integral is an Itô stochastic integral
against the noise W it is defined in the
mean-square (L²) sense, not pathwise.
The solution V(x, t) is not a single function.
It is a measurable map:
V : Ω × [0,T] × ℝ → ℝ
where Ω is the probability space carrying W.
For each ω ∈ Ω (each scenario), V(·,·,ω) is
one realisation one possible price surface.
4) SPDE models:
Heston (1993) volatility with its own gravity :
The core insight of Heston is that volatility is not a number .It is a mean-reverting process. It wanders, but it has a gravitational pull back toward a long-run level. The asset price and its volatility are coupled and correlated, which is what generates the asymmetric smile , the skew .
SABR (2002) the rates trader’s model
SABR was built for interest rate derivatives, where the forward rate and its volatility move together in a way that produces pronounced smiles and skews across expiries. Its dominant advantage is that it has an approximate closed-form solution for implied volatility meaning traders can calibrate it instantly and use it to quote options across the entire strike surface without running a Monte Carlo simulation.
REAL MODELS IN SPDE TERRITORY
MODEL 1 :HESTON (1993)
Stochastic Volatility with Mean Reversion
Two coupled SDEs drive the system:
dS_t = μ S_t dt + √v_t · S_t dW_t^S
dv_t = κ(θ − v_t) dt + ξ √v_t dW_t^v
with the correlation structure:
dW_t^S · dW_t^v = ρ dt, ρ ∈ [−1, 1]
Parameters:
S_t asset price at time t
v_t instantaneous variance (v_t = σ_t², stochastic)
μ drift of the asset price
κ mean-reversion speed of variance
θ long-run mean of variance (v_t → θ as t → ∞)
ξ vol-of-vol : volatility of the variance process
ρ correlation between asset and variance shocks
W_t^S Brownian motion driving the asset price
W_t^v Brownian motion driving the variance
The Feller condition for v_t > 0 almost surely:
2κθ > ξ²
If this condition is violated, the variance process can
reach zero, which collapses the model to a degenerate state.
- - - - - - - - - - - - - - - - - - - - - - - - - - - - -
THE HESTON PDE
- - - - - - - - - - - - - - - - - - - - - - - - - - - - -
By Itô's lemma applied to V(S, v, t), the no-arbitrage
condition produces a 2D PDE in (S, v, t):
∂V/∂t
+ (1/2)vS² · ∂²V/∂S²
+ ρξvS · ∂²V/∂S∂v
+ (1/2)ξ²v · ∂²V/∂v²
+ rS · ∂V/∂S
+ [κ(θ−v) − λv] · ∂V/∂v
− rV
= 0
where λ is the market price of volatility risk.
This is not the heat equation.
It is a degenerate parabolic PDE : the diffusion matrix:
| vS² ρξvS |
A = | |
| ρξvS ξ²v |
is positive semi-definite but not uniformly elliptic.
It degenerates when v → 0, which is precisely the Feller
boundary and the source of most of the numerical difficulty.
- - - - - - - - - - - - - - - - - - - - - - - - - - - - -
WHY THE SMILE EMERGES
- - - - - - - - - - - - - - - - - - - - - - - - - - - - -
ρ < 0 (negative correlation, typical for equities):
When S falls, v tends to rise.
This produces left skew — higher implied vol for
low strikes than high strikes.
Matches the persistent equity volatility skew.
ξ > 0 (vol-of-vol is nonzero):
The variance process itself has dispersion.
This fattens the tails of the return distribution
symmetrically, generating smile curvature.
Together ρ and ξ give full control over the shape of
the implied volatility surface across strikes and expiries.
- - - - - - - - - - - - - - - - - - - - - - - - - - - - -
SOLUTION METHOD
- - - - - - - - - - - - - - - - - - - - - - - - - - - - -
No elementary closed form exists for general payoffs.
Heston (1993) derived a semi-analytical solution for
European calls via the characteristic function φ(u):
φ(u; S, v, t) = exp( C(u,τ) + D(u,τ)·v + iu·ln(S) )
where τ = T − t, and C(u,τ), D(u,τ) solve Riccati ODEs:
D(u,τ) = (κ − ρξiu − d) / ξ²
· (1 − e^(−dτ)) / (1 − g·e^(−dτ))
C(u,τ) = rτiu + (κθ/ξ²)·[(κ−ρξiu−d)τ
− 2 ln((1 − g·e^(−dτ)) / (1−g))]
d = √[ (ρξiu − κ)² + ξ²(iu + u²) ]
g = (κ − ρξiu − d) / (κ − ρξiu + d)
The call price is then recovered by Fourier inversion:
C(S, v, t) = S · P₁ − K·e^(−rτ) · P₂
P_j = (1/2) + (1/π) ∫₀^∞ Re[ e^(−iu ln K)
· φ_j(u) / (iu) ] du
where φ₁ and φ₂ are the two characteristic functions
corresponding to the two probability terms (analogous
to N(d₁) and N(d₂) in Black–Scholes).
MODEL 2 : SABR (2002)
Stochastic Alpha Beta Rho
SABR models the forward price F_t and its volatility α_t
as jointly stochastic, with no mean reversion on α_t:
dF_t = α_t · F_t^β · dW_t^F
dα_t = ν · α_t · dW_t^α
with the correlation structure:
dW_t^F · dW_t^α = ρ dt, ρ ∈ [−1, 1]
Parameters:
F_t forward price (or forward rate) at time t
α_t stochastic volatility process (α_0 = α, initial vol)
β elasticity parameter β ∈ [0, 1]
ν vol-of-vol (ν > 0)
ρ correlation between F and α shocks
Note: α_t follows a geometric Brownian motion with no drift.
It has no mean reversion — unlike Heston's variance process.
This means α_t can drift arbitrarily, which is realistic for
interest rate volatility over long horizons.
- - - - - - - - - - - - - - - - - - - - - - - - - - - - -
THE β PARAMETER : CEV BACKBONE
- - - - - - - - - - - - - - - - - - - - - - - - - - - - -
β controls how the local volatility scales with F_t:
β = 1 → dF_t = α_t F_t dW_t^F
Log-normal backbone (Black's model)
Constant percentage vol regardless of level
β = 0 → dF_t = α_t dW_t^F
Normal backbone (Bachelier model)
Constant absolute vol regardless of level
β = 1/2 → dF_t = α_t √F_t dW_t^F
CIR-type backbone
Vol scales with square root of forward
In interest rate markets, β = 0.5 or β = 1 are most common.
In FX markets, β = 1 is the standard.
- - - - - - - - - - - - - - - - - - - - - - - - - - - - -
THE HAGAN ET AL. IMPLIED VOLATILITY FORMULA
- - - - - - - - - - - - - - - - - - - - - - - - - - - - -
The key result of SABR is an analytic approximation for
the implied Black volatility σ_B(F, K, T) as a function
of strike K and expiry T (Hagan et al., 2002):
σ_B(F, K, T) ≈
α · z
─────────────────────────────────────────────────── ·
(FK)^((1−β)/2) · χ(z) · [1 + A₁T + A₂T²/...]
[1 + ( (1−β)²/24 · α²/(FK)^(1−β)
+ ρβνα/(4(FK)^((1−β)/2))
+ (2−3ρ²)/24 · ν² ) · T ]
where:
z = (ν/α) · (FK)^((1−β)/2) · ln(F/K)
χ(z) = ln[ (√(1−2ρz+z²) + z − ρ) / (1 − ρ) ]
For ATM (F ≈ K), the formula simplifies to:
σ_ATM ≈ α / F^(1−β) ·
[ 1 + ( (1−β)²/24 · α²/F^(2−2β)
+ ρβνα/(4F^(1−β))
+ (2−3ρ²)/24 · ν² ) · T ]
- - - - - - - - - - - - - - - - - - - - - - - - - - - - -
WHY THE SMILE EMERGES IN SABR
- - - - - - - - - - - - - - - - - - - - - - - - - - - - -
ρ ≠ 0 controls skew:
ρ < 0 → downward skew (higher vol for low strikes)
ρ > 0 → upward skew (higher vol for high strikes)
ν > 0 controls smile curvature:
Higher ν → fatter tails → more pronounced smile
ν = 0 → flat vol surface (SABR collapses to CEV)
β controls the backbone:
Determines how implied vol shifts as the forward
moves critical for hedging in rate markets where
forwards can move hundreds of basis points.
Both models live in SPDE territory because in both cases
the diffusion coefficient of the pricing equation is itself
stochastic it cannot be written as a fixed function of
(x, t) alone, and the solution is a random field rather
than a deterministic surface.
5) The Cost of Structuring Uncertainty:
Black-Scholes had one number to calibrate σ. You could imply it from a single option price in milliseconds. SPDEs ask you to calibrate entire processes, correlation structures, and noise coefficients simultaneously, across a surface of market prices that is itself moving. The richness is real. So is the cost.
- Numerical Solutions are Expensive :A classical PDE lives in two dimensions :price and time. You discretise a grid, run a finite difference scheme, and you’re done. An SPDE solution is a random field. To approximate it you need Monte Carlo simulation over entire realisations of the driving noise, which means running thousands of paths, each of which is itself a function of space and time. Multilevel Monte Carlo methods exist to reduce the variance of these estimates by simulating at multiple levels of resolution and combining them but even that is orders of magnitude more expensive than solving a classical PDE on a grid. In a trading environment where Greeks need to be recalculated in real time as markets move, that gap matters enormously.
- Calibration is Harder Heston has five parameters: κ, θ, ξ, ρ, v₀. Each one needs to be estimated from market prices, which means solving a nonlinear least-squares problem over an implied volatility surface. The objective function is non-convex. Multiple parameter sets can fit the surface equally well and produce completely different hedges and dynamics. SABR has four parameters but the Hagan approximation breaks down for long expiries and extreme strikes, meaning you’re calibrating a formula that you know is wrong in the tails which is exactly where the risk lives during stress events.
In conclusion:
Markets are not smooth. They never were. Black-Scholes gave us a foundation , elegant, useful, and wrong in exactly the ways that matter most when things go bad and rough . SPDEs are the honest version of the same question: what is the price of uncertainty when uncertainty itself is uncertain?
The cost is real computationally, analytically, and operationally. But the direction is correct. Every serious model used in production today, from Heston to SABR to stochastic local volatility, is quietly living in SPDE territory whether it calls itself that or not. The field has already moved. The mathematics just gives it a name.
!!! Happy Reading
