Finite Difference Methods for Options

The finite difference method (FDM) solves the option pricing problem by discretizing the Black-Scholes partial differential equation (PDE) on a time-and-price grid. Rather than computing option values via closed-form formulas or lattice recursion, finite differences replace continuous derivatives with discrete differences, converting the PDE into a set of algebraic equations that are solved step-by-step in time. Three main schemes—explicit, implicit, and Crank-Nicolson—trade off speed, stability, and accuracy.

The Black-Scholes PDE and its discretization

The Black-Scholes PDE states that an option value V(S, t) satisfies:

∂V/∂t + (1/2)σ²S²∂²V/∂S² + r(S∂V/∂S − V) = 0

where σ is the volatility, r is the risk-free interest rate, and S is the stock price. This equation holds for any option (European, American, path-dependent) under the assumptions of no arbitrage and no transaction costs.

Finite difference methods discretize this equation by replacing the continuous space (S, t) with a grid: S_i = i·Δs (for i = 0, 1, …, I) and t_j = j·Δt (for j = 0, 1, …, J). The option value at grid point (i, j) is denoted V_{i,j}.

The partial derivatives are approximated as finite differences. For example:

• ∂V/∂t ≈ (V_{i,j+1} − V_{i,j}) / Δt (forward difference in time)
• ∂V/∂S ≈ (V_{i+1,j} − V_{i−1,j}) / (2Δs) (central difference in space)
• ∂²V/∂S² ≈ (V_{i+1,j} − 2V_{i,j} + V_{i−1,j}) / (Δs)²

Substituting these into the PDE yields an algebraic equation relating V_{i,j} to its neighbors. Depending on how the time derivative is handled, different schemes emerge.

The explicit scheme: fast but fragile

The explicit scheme uses a forward difference in time. Rearranging the discretized PDE, the option value at the next time step is:

V_{i,j+1} = a·V_{i−1,j} + b·V_{i,j} + c·V_{i+1,j}

where a, b, c are coefficients that depend on σ, r, Δs, and Δt.

This formula is explicit: V_{i,j+1} is computed directly from the previous time step’s values. Starting from the final condition (V(S, T) = payoff at maturity T), the algorithm works backward in time, computing each V_{i,j} from V_{i,j+1}.

The explicit scheme is fast—each time step requires only arithmetic, no matrix inversion. However, it is conditionally stable: for accuracy and to avoid oscillation, Δt must satisfy:

Δt ≤ (Δs)² / (σ² × S_max²)

roughly. If you choose Δs = 0.1 (fine spatial grid) and σ = 0.2, this constraint might require Δt ≤ 0.00005, forcing many time steps and offsetting the speed gain. Stability failure produces nonsensical (oscillating or exploding) option values.

For this reason, explicit schemes are rarely used in production. They are pedagogical—they show the core idea clearly—but unreliable for real trading.

The implicit scheme: stable, matrix-heavy

The implicit scheme uses a backward difference in time. Rearranging yields:

a·V_{i−1,j} + b·V_{i,j} + c·V_{i+1,j} = V_{i,j+1}

Now the option values at time t_j on the left (unknown) depend on values at t_{j+1} on the right (known). At each time step, you must solve a tridiagonal linear system—efficient with banded-matrix solvers—to find all V_{i,j} simultaneously.

The implicit scheme is unconditionally stable: no constraint on Δt relative to Δs. You can choose time and space steps independently. This flexibility is invaluable.

However, the implicit scheme is only first-order accurate in time. If you reduce Δt by a factor of 10 to improve accuracy, you must solve 10 times as many linear systems. For a fine grid (large I), this becomes expensive.

Crank-Nicolson: the production workhorse

The Crank-Nicolson (CN) scheme averages the explicit and implicit schemes. It uses a central difference in time, effectively evaluating the PDE at the midpoint between t_j and t_{j+1}:

a·V_{i−1,j} + b·V_{i,j} + c·V_{i+1,j} = a’·V_{i−1,j+1} + b’·V_{i,j+1} + c’·V_{i+1,j+1}

This rearranges to:

(b’ − a’·V_{i−1,j+1} − c’·V_{i+1,j+1}) − [(b − a·V_{i−1,j} − c·V_{i+1,j})] = 0

Again, a tridiagonal system, but now Crank-Nicolson achieves second-order accuracy in both time and space. This is a significant gain: the discretization error shrinks as O((Δt)² + (Δs)²), compared to O(Δt + (Δs)²) for implicit.

The CN scheme is unconditionally stable (like implicit) and far more accurate (like explicit, but without the stability headache). It is the industry standard for option pricing.

Handling American options and early exercise

An American option value must satisfy:

V(S, t) ≥ intrinsic value

at every point in time and space. If the equation (V from the PDE) gives a value below intrinsic, the option holder exercises immediately.

In finite difference form, at each grid point (i, j) and each time step, after solving the linear system, you apply a free-boundary constraint:

V_{i,j} = max(V_{i,j}, intrinsic_{i,j})

This is simple to implement and automatically enforces early exercise. For American calls on non-dividend stocks, the optimal exercise boundary (the stock price at which to exercise) emerges from this max operation.

Implicit and Crank-Nicolson schemes handle this naturally. Explicit schemes can also include the constraint, though the stability condition becomes even more restrictive.

Boundary conditions

Finite difference grids are bounded: S ranges from 0 to S_max and t from 0 to maturity T. The PDE solution requires boundary conditions:

• At S = 0: most equities approach zero asymptotically, and an option value on a dead stock depends only on type. For a call, V = 0. For a put, V = strike × exp(−r·τ) where τ is time to maturity.
• At S = S_max: far out of the money, options approach their intrinsic value. Often, V ≈ S − strike (for a call).
• At t = T: the final payoff. V(S, T) = max(S − K, 0) for a call.

Choosing S_max appropriately is crucial. Too small, and the boundary affects the interior solution. Too large, and the grid becomes coarse and inaccurate. A rule of thumb: S_max = 3–4 times the strike.

Convergence and error analysis

For Crank-Nicolson, the discretization error decreases as O((Δt)² + (Δs)²). In practice, doubling the grid density (halving both Δt and Δs) reduces error by a factor of ~4 in each direction, or ~16 overall.

Typical production grids use:

• I = 100–500 stock-price levels
• J = 100–500 time steps
• Runtime: milliseconds on a modern CPU

For a book of 10,000 options, a per-contract FDM evaluation takes ~0.1 ms, enabling fast risk updates.

When to use finite differences versus alternatives

Finite differences excel for:

• American options, especially with early-exercise decisions.
• Local volatility (e.g., CEV model), where volatility varies with stock price.
• Exotic payoffs (e.g., knock-out barriers, averaging).
• Path-dependent features that require state variables.

Lattice methods (Binomial, Trinomial) are intuitive, especially for teaching, and converge quickly with a smooth lattice.

Fourier methods (FFT) are fastest for European options on simple underlying assets, but cannot handle early exercise.

Monte Carlo is flexible for multi-asset and path-dependent options but is slow for vanilla European options.

For production equity-option trading systems, Crank-Nicolson finite differences are often the first choice.

See also