American Option Pricing Approaches

Pricing American options means solving a problem European options don't have: at every moment before expiration, the holder can exercise. That early exercise right creates a free boundary — a moving line in price-time space that separates "exercise now" from "keep holding." Every pricing method is fundamentally an approach to finding that boundary. Get it wrong and you either overpay for optionality you won't use or underprice the right to act early.

The point is: Only about 7% of options are actually exercised (OCC, 2024), and most of those happen at or near expiration. But the right to exercise early still has value — empirically 0.5–3% of option price for near-the-money equity options, and up to 14.4% for deep out-of-the-money index options (Li et al., 2024, Journal of Futures Markets). Your pricing model must capture that value accurately, even though it rarely gets realized.

The Early Exercise Boundary (What You're Actually Solving For)

The optimal exercise boundary S*(t) divides every point in (stock price, time) space into two regions:

For American puts:

• If S < S*(t): exercise immediately — the interest earned on the strike cash exceeds the remaining time value
• If S >= S*(t): hold — there's still enough optionality to justify waiting

For American calls on dividend-paying stocks:

• Exercise becomes optimal only at ex-dividend dates, when remaining time value drops below the dividend amount
• Calls on non-dividend stocks are never optimally exercised early — this is a fundamental result, not an approximation

Why this matters for pricing: A model that mislocates the exercise boundary by even a few dollars in stock price will systematically misprice the option. The boundary shifts discontinuously at ex-dividend dates, which is why discrete dividend handling matters far more than most practitioners realize. Systems that collapse discrete dividends into a continuous yield "induce spurious correlation between prediction error and moneyness, probability of early exercise, time to expiration, and dividend yield" (Damodaran, NYU).

Three Methods, One Problem

The practical guidance is simple: dimensionality is the dominant selection criterion. For single-asset vanilla Americans, binomial trees or finite differences win on both accuracy and speed. Once you have two or more underlying assets, Longstaff-Schwartz is the only viable choice.

Binomial Trees (The Workhorse)

At each node, the algorithm compares two values:

continuation_value = discount × (p × V_up + (1-p) × V_down)
intrinsic_value = max(K - S, 0)   // for a put
V_node = max(continuation_value, intrinsic_value)

If intrinsic beats continuation, that node sits inside the exercise region.

What practitioners actually care about:

• 200 steps is the standard for production pricing of single-asset Americans — a 200-step CRR tree returns prices in under 5 milliseconds for standard parameters
• CRR trees oscillate. Even vs. odd step counts give systematically different prices depending on whether the strike falls on a node. The Leisen-Reimer tree and Tian third-moment matching tree minimize this problem and are preferred for production
• Richardson extrapolation accelerates convergence but works best with Black-Scholes smoothing at terminal nodes. Without smoothing, unadapted CRR trees converge at first order; with it, you can achieve order 3/2 (Joshi, 2009)
• Discrete dividends are handled naturally — just reduce the stock price at the ex-date node. This is the binomial tree's single biggest practical advantage over Black-Scholes approximations

The test: Run your tree with 199 and 200 steps. If the prices differ by more than a few cents, your implementation isn't smoothed properly.

Finite Difference Methods (The Accuracy Standard)

Finite difference discretizes the Black-Scholes PDE on a grid and solves backward in time. At each grid point, the American constraint forces the solution to be at least the intrinsic value:

V_j^n = max(V_j^n_from_PDE, intrinsic_j)

Production parameters:

• Research benchmarks use 512 space steps x 1,024 time steps; production systems typically run 500 spatial x 100+ temporal points
• Non-uniform grid spacing (e.g., hyperbolic sine transformation concentrating points near the strike) dramatically improves accuracy per grid point — this is standard practice, not optional refinement
• Crank-Nicolson is the classic time-stepping scheme, but the TR-BDF2 scheme (Le Floc'h) handles the American free boundary with better stability
• Finite differences with a 400x400 grid deliver accuracy that binomial trees need 500+ steps to match

Why this matters: FD methods give you Greeks directly from the grid (delta and gamma from spatial differences, theta from time differences) without bumping. For end-of-day risk reports, this is a significant operational advantage.

Longstaff-Schwartz (When Dimensions Multiply)

The algorithm:

1. Simulate paths: Generate N paths of the underlying(s) from t=0 to T
2. Terminal payoffs: At T, compute payoff for each path
3. Backward recursion: At each exercise date before T:
   • Identify in-the-money paths
   • Regress discounted future cash flows on basis functions of current stock price (typically 1, S, S² — Laguerre polynomials are standard)
   • Exercise if intrinsic exceeds the fitted continuation value
4. Average discounted payoffs across all paths

Production reality:

• The original Longstaff-Schwartz paper (2001) used 50,000 paths (25,000 + 25,000 antithetic) with 3 Laguerre basis functions
• Production implementations run 50,000–100,000 paths — more paths reduce noise but hit diminishing returns
• The regression is the fragile part. Basis function selection matters: too few and you miss the exercise boundary's curvature; too many and you overfit to noise. Three to five functions is the standard
• For vanilla single-asset Americans, LS-MC is not competitive with trees or FD on speed. Its advantage is exclusively dimensional: it scales linearly with the number of underlying assets

What this means in practice: Longstaff-Schwartz exists because trees and FD methods are exponential in dimensions. A basket option on 5 stocks is trivial for LS-MC (just add columns to the simulation matrix) and effectively impossible for a 5-dimensional binomial tree.

The State of the Art (What's Actually Changing)

Analytic approximations have won for vanilla Americans. The Andersen-Lake-Offengenden (2015) spectral collocation method, implemented in QuantLib as QdFpAmericanEngine, prices at ~100,000 options per second per CPU core with accuracy to 10–11 significant digits. Per HPC-QuantLib benchmarks on a portfolio of 6,000 options across varying strikes, maturities, rates, and vols: "PDEs cannot compete with this algorithm, nor can tree-based algorithms."

The practical implication: For single-asset vanilla Americans in a Black-Scholes world, the pricing problem is solved. Trees and FD still matter for:

• Barrier options and exotic payoffs (FD handles boundaries naturally)
• Teaching and intuition-building (trees)
• Stochastic volatility (FD Heston engine, e.g., QuantLib's FdHestonVanillaEngine using a 3D grid: time × price × vol)

Neural network pricing is emerging for production use. Anderson & Ulrych (2023) trained feed-forward networks over Heston model parameters to produce near-instantaneous American prices after a GPU training phase. Reppen, Soner & Tissot-Daguette (2025, Mathematical Finance) parameterized the stopping boundary itself as a neural network — learning the optimal exercise surface directly. These approaches trade offline training time for microsecond inference, which matters for high-frequency market making where FPGA-based systems target sub-microsecond tick-to-trade latency.

Validation (Don't Ship Without These)

Mandatory checks before production:

• American >= European: For identical terms, the American price must always equal or exceed the European. If your model ever violates this, you have a bug — no exceptions
• Zero-rate convergence: At r = 0%, American puts should equal European puts (no interest incentive to exercise early). This catches boundary condition errors
• Convergence test: Increase steps/paths/grid by 2x. If the price moves by more than your tolerance, you need more resolution
• Early exercise boundary shape: The put boundary S*(t) must be monotonically increasing, reaching the strike K at expiration. Any non-monotonic kink signals a numerical artifact

High-impact refinements:

• Compare your implementation against QuantLib's QdFpAmericanEngine on benchmark cases — match within 0.01% for normal parameters, 0.1% for stress parameters
• Stress test at sigma = 200%, at zero rates, and with large discrete dividends
• Verify Greeks are smooth across the exercise boundary (discontinuous delta is a red flag in FD implementations)

Checklist: Choosing Your Method

Essential (Start Here)

• Know your dimensionality: 1 asset → tree or FD; 2+ assets → Longstaff-Schwartz
• For vanilla single-asset Americans, use an analytic approximation (Andersen-Lake-Offengenden) unless you need Greeks from the grid
• Handle discrete dividends explicitly — continuous yield approximations misprice early exercise boundaries

High-Impact (Production Hardening)

• Implement Richardson extrapolation or use Leisen-Reimer trees to eliminate oscillatory convergence
• For FD, use non-uniform grids concentrated near the strike and TR-BDF2 time stepping
• Validate against QuantLib benchmarks before going live

Advanced (For Pricing Libraries)

• Explore neural network approximators for real-time pricing under stochastic volatility
• Implement Heston FD on a 3D grid (time × price × variance) for vol-surface-consistent American prices
• Profile LS-MC basis function sensitivity — run 3, 4, and 5 Laguerre polynomials and measure exercise boundary stability

---

The bottom line: American option pricing is a solved problem for vanilla single-asset contracts — the Andersen-Lake-Offengenden method delivers 11-digit accuracy at 100,000 prices per second. The complexity lives in multi-asset problems (where Longstaff-Schwartz scales linearly), stochastic volatility (where FD Heston operates on 3D grids), and the emerging frontier of neural network approximators that trade training time for microsecond inference. The question isn't which method is "best" — it's which problem you're actually solving. A single-stock American put and a basket option on five correlated equities require fundamentally different tools, and using the wrong one wastes either accuracy or compute.

Citation: Longstaff, F. & Schwartz, E. (2001). Valuing American Options by Simulation. Review of Financial Studies, 14(1), 113-147. Andersen, L., Lake, M., & Offengenden, D. (2015). High-Performance American Option Pricing. Journal of Computational Finance. Li, Z. et al. (2024). An Empirical Study on the Early Exercise Premium. Journal of Futures Markets, 44(7), 1117-1153.