How Longstaff–Schwartz Least-Squares Monte Carlo Algorithms Price Early Exercise: Simulated Paths, Continuation Regressions, Exercise Rules, Bias and Dual Bounds

Reader question: Monte Carlo simulation naturally moves forward through possible future paths. American-style exercise decisions work backward because today’s choice depends on the value of continuing into an uncertain future. How can one algorithm combine those opposite directions?

The Longstaff–Schwartz least-squares Monte Carlo method does it by separating the problem into two passes. First, it simulates many state paths forward under the pricing model. Then it walks backward across exercise dates and uses regression to estimate the conditional continuation value: the expected discounted value of keeping the option alive instead of exercising now.

The result is not merely a price. It is an approximate stopping rule: at each exercise date and simulated state, compare immediate exercise value with estimated continuation value and choose the larger.

What this page owns — and what it does not

This page owns the regression-based optimal-stopping transformation:

simulated risk-neutral paths + exercise payoff + basis functions → approximate continuation values → exercise policy → option value.

It does not replace binomial-tree American-option pricing, which owns lattice backward induction; general Monte Carlo pricing, which owns simulation mechanics; or Heston calibration, which determines parameters for a stochastic-volatility model that could later feed the simulation.

This is derivative-pricing mathematics, not an options strategy and not personalized financial advice.

The optimal-stopping equation

At an exercise date t, let the immediate exercise payoff be:

g_t(X_t).

Let the continuation value be the conditional expectation of future optimal discounted cash flow:

C_t(X_t) = E_t[D(t,t+1)V_t+1].

The option value satisfies the Bellman-style recursion:

V_t = max(g_t, C_t).

At maturity, there is no continuation decision:

V_T = g_T.

The hard quantity is C_t. In a tree it is obtained from child nodes. In high-dimensional Monte Carlo there are no reusable child nodes, so Longstaff and Schwartz estimate it statistically.

Step 1: simulate paths forward

Choose a risk-neutral model for the underlying state process and simulate paths:

X₀^(m), X₁^(m), …, X_T^(m),

for paths m = 1,…,M.

For a simple equity option, X may be the stock price. In a multifactor interest-rate derivative, it can include several rates or curve factors. For a path-dependent contract, the state must also include sufficient path information such as an average, running maximum or accumulated coupon state.

This state sufficiency is a major weak link. If the regression omits a variable that affects future value, continuation estimates can be biased even with infinite simulation paths.

Step 2: start at maturity

For every path, calculate the terminal payoff.

For an American/Bermudan put with strike K:

g_T = max(K − S_T, 0).

Those pathwise terminal cash flows become the first known future cash flows when the algorithm steps backward to the prior exercise date.

Step 3: build the regression target

At an earlier exercise date t, suppose a path has a future cash flow Y at a later date under the decisions already determined farther down the path.

Discount that future cash flow back to t:

Y_t = D(t,τ) × CashFlow_τ.

This realized discounted future cash flow is a noisy sample of the conditional continuation value.

The regression asks:

Y_t^(m) ≈ β₀φ₀(X_t^(m)) + β₁φ₁(X_t^(m)) + … + β_Kφ_K(X_t^(m)).

The fitted value is:

Ĉ_t(X) = Σ β̂_kφ_k(X).

Why least squares appears

The true continuation value is a conditional expectation. Conditional expectation is also the mean-square-optimal predictor of a future random variable given the current information set.

By projecting realized discounted future cash flows onto a chosen family of basis functions, least squares approximates that conditional expectation inside the function space spanned by the basis.

This is the mathematical bridge between Monte Carlo and dynamic programming.

Why the original method focuses on in-the-money paths

If immediate exercise payoff is zero, the holder will normally not exercise a nonnegative-value option merely to receive zero while continuation remains available.

Longstaff–Schwartz therefore commonly estimates the exercise regression using paths that are in the money at the exercise date.

This focuses scarce regression capacity near states where exercise is actually a live decision.

It is not a universal command that every modern variant must exclude all out-of-the-money paths. It is a design choice grounded in the exercise decision and the original algorithm’s efficiency.

Step 4: compare exercise with continuation

For an in-the-money path:

Exercise now if g_t(X_t) ≥ Ĉ_t(X_t).

Otherwise continue.

If exercise occurs, all later cash flows on that path are deleted because the contract terminates at the chosen stopping time. If continuation occurs, the later selected cash flow remains.

The algorithm then moves to the previous exercise date and repeats.

A tiny American-put intuition

Suppose at a candidate exercise date:

• strike = 100;
• current simulated stock price = 90;
• immediate exercise payoff = 10;
• regression-estimated continuation value = 8.40.

The rule exercises because 10 > 8.40.

If the estimated continuation value were 12.10, the option would continue because the expected discounted value of waiting exceeds immediate exercise.

The hard part is not the comparison. It is producing a continuation estimate reliable enough near the exercise boundary.

Basis functions define what shapes the regression can learn

For a one-dimensional stock price, a simple basis can be:

{1, S, S²}.

Longstaff and Schwartz also used families such as Laguerre polynomials in examples.

In a multifactor problem, basis functions can include cross-products:

{1, X₁, X₂, X₁², X₂², X₁X₂, …}.

Too few basis functions underfit continuation value. Too many relative to the number of effective regression paths create unstable coefficients and overfitting.

The regressors-to-paths ratio matters

A 2024 study by Liu, Woo and Choi on leave-one-out least-squares Monte Carlo showed that look-ahead bias is linked to the ratio between the number of regressors and simulation paths.

This gives a practical diagnostic: a regression with 40 basis terms and only 60 relevant in-the-money paths is not “more sophisticated” than one with six robust terms. It may be almost fitting noise path by path.

Scaling and conditioning

Raw powers of a large state variable can create severe numerical conditioning problems.

If S is near 10,000, then S² is near 100 million. A regression matrix combining 1, 10,000 and 100,000,000 can be ill-conditioned.

Safer approaches include:

• normalizing state variables;
• using orthogonal polynomial families;
• QR or SVD-based least squares rather than naive normal equations;
• dropping nearly collinear basis terms.

A price can look stable to cents while regression coefficients are numerically explosive. Condition-number diagnostics matter.

Look-ahead bias: the path helps fit its own exercise rule

A subtle problem appears if the same simulated paths are used both to estimate continuation and to value the resulting exercise policy.

The future realized cash flow from a path helps determine the regression coefficient, and that fitted regression is then used to decide whether the same path should exercise. Information from that path’s future has leaked into its current stopping decision through estimation.

This can create look-ahead bias and can overvalue the option relative to a truly implementable stopping policy.

Training paths and valuation paths

A clean control is:

1. simulate a training sample;
2. fit the exercise policy backward on that sample;
3. freeze the policy;
4. simulate an independent valuation sample;
5. apply the frozen policy without refitting.

The resulting out-of-sample value is a lower bound on the true Bermudan/American value under the pricing model, apart from statistical estimation error in the sample mean, because a fixed suboptimal stopping rule cannot be worth more than the optimal stopping rule.

Leave-one-out LSMC

The 2024 leave-one-out LSMC method targets the same leakage more efficiently. For each training observation, it estimates the continuation prediction as though that observation had not helped fit its own regression.

This reduces the self-influence that produces look-ahead bias without requiring a completely separate second simulation just to remove that source of bias.

The broader lesson is not that one modern correction is mandatory. It is that policy training and policy valuation must not silently reuse future information in a way an actual exercise decision could not.

Lower bounds and dual upper bounds

An implementable stopping policy gives a primal lower bound.

Rogers and, independently, later primal-dual methods show that American-option optimal stopping has a dual representation involving martingales. Andersen and Broadie developed a practical simulation algorithm that combines a lower-bound exercise policy with an upper-bound construction.

If the lower and upper bounds are close, the exercise policy is strongly validated:

Lower bound ≤ true option value ≤ upper bound.

A wide duality gap is evidence that continuation estimates or the stopping policy remain weak.

Time discretization creates another approximation layer

A continuously exercisable American option is approximated in Monte Carlo by a finite set of candidate exercise dates unless a specialized continuous-time treatment is used.

With monthly exercise dates, the algorithm prices a Bermudan approximation, not literally continuous exercise.

Increasing exercise-date frequency should generally move the value toward the continuous American limit, but it also makes regression harder because there are more backward steps and often fewer informative paths near each boundary.

Three important error sources

Simulation error: finite paths create Monte Carlo noise.

Regression approximation/estimation error: the basis family and finite sample approximate the conditional continuation function imperfectly.

Exercise-grid error: a finite set of dates approximates continuous exercise when the contract is truly American.

A reported standard error for the final sample mean does not automatically capture all three.

High dimensionality: the reason LSMC is attractive

Finite-difference PDE methods suffer strongly as the number of state variables rises because the grid grows exponentially.

Simulation can handle many stochastic factors more naturally. Longstaff and Schwartz highlighted path-dependent and multifactor examples, including high-dimensional term-structure settings.

But the curse of dimensionality has not disappeared. It has moved into continuation-value regression: the basis set needed to approximate a function of many state variables can still grow explosively.

Path dependence requires state augmentation

Suppose an option payoff depends on the running average price. Two paths can have the same current stock price but different averages and therefore different future exercise economics.

If the regression uses only current stock price, it treats those states as identical.

The state vector must include sufficient information such as:

X_t = (S_t, Average_t).

For callable loans, storage contracts or interest-rate products, the required state can include accrued quantities, remaining rights, curve factors or other path variables.

Inputs and outputs

A robust LSMC engine can require:

• underlying risk-neutral process and calibrated parameters;
• exercise dates;
• payoff function;
• discounting model;
• state variables and path-dependent state updates;
• number of simulation paths;
• time grid;
• basis functions and scaling rules;
• ITM-path selection policy;
• random-number/variance-reduction settings;
• training-versus-valuation sample policy.

Outputs should include:

• option value and sampling uncertainty;
• exercise policy or boundary diagnostics;
• regression coefficients by exercise date;
• number of effective regression paths;
• condition numbers or rank diagnostics;
• out-of-sample lower bound;
• dual upper bound where available;
• convergence tests across paths, basis and time grids.

Evidence polarity: what supports confidence?

Evidence for a good LSMC implementation includes convergence toward a trusted binomial/PDE benchmark in low-dimensional cases, stable prices as path counts increase, sensible exercise boundaries, good out-of-sample policy value, a small primal-dual gap, well-conditioned regressions and monotonicity with respect to obvious contract parameters.

Evidence against confidence includes large changes when one basis term is added, exercise decisions in economically impossible regions, a value below intrinsic value, an in-sample price materially above an independent policy valuation, singular regression matrices, very few ITM paths, or a wide dual upper/lower interval.

Counterexample: more basis functions can make the price worse

If a low-order polynomial captures the relevant continuation shape, adding many high-order terms can fit simulation noise.

The in-sample regression error falls, but the stopping rule becomes less stable out of sample.

Model complexity should therefore be judged by policy performance and convergence, not merely by regression R².

Counterexample: a high R² does not prove a correct exercise boundary

Most discounted future cash-flow variation can occur far away from the exact exercise/continue boundary. A regression can fit overall continuation cash flows well while making enough local errors near the boundary to distort exercise decisions materially.

Boundary-focused diagnostics are more informative than global fit alone.

Counterexample: an American call on a non-dividend-paying stock

Under the standard frictionless Black–Scholes assumptions with nonnegative interest rates and no dividends, early exercise of a plain American call is not optimal.

An LSMC implementation that repeatedly exercises such a call early under those assumptions has revealed a regression, discounting or simulation problem.

This makes the contract a powerful unit test because the correct stopping policy is known qualitatively.

Counterexample: American value below European value

An American option includes every European exercise strategy plus additional early-exercise rights.

Therefore:

V_American ≥ V_European.

If an LSMC estimate is materially below a reliable European benchmark beyond Monte Carlo uncertainty, the stopping policy or estimator is underperforming badly.

Weak links in implementation

Wrong probability measure. Simulating physical-return drift instead of the pricing measure misprices continuation.

Discounting mismatch. Future cash flows are regressed without consistent numeraire/discount treatment.

State omission. Path-dependent information affecting continuation is missing.

In-sample self-use. The path helps fit its own stopping decision and creates look-ahead bias.

Ill-conditioned basis. Normal equations amplify numerical error.

Too few ITM paths. Near maturity or for deep OTM regions, regression becomes unstable.

Exercise-grid mismatch. A monthly Bermudan engine is compared directly with a continuously exercisable American benchmark.

Random-seed dependence. A price that changes materially across seeds lacks enough effective simulation.

Diagnostics: how to test the algorithm

• European test: disable early exercise and compare with a known European price.
• American-call no-dividend test: confirm the policy almost never exercises early under the classical assumptions.
• binomial benchmark: compare one-factor American puts with the CRR lattice.
• path-doubling test: increase simulation paths and monitor value plus standard error.
• basis ladder: increase basis richness gradually and require out-of-sample stability.
• grid refinement: add exercise dates and monitor convergence.
• train/test split: fit a policy on one sample and value it on an independent sample.
• leave-one-out test: compare conventional in-sample LSM with a look-ahead-bias correction.
• dual bound test: compute an upper bound where practical and inspect the gap.
• boundary plot: visualize exercise/continue states and look for non-economic discontinuities.

What would falsify confidence?

Confidence should be withdrawn if independent valuation paths produce a materially lower value than the reported in-sample estimate; if the method fails known qualitative exercise tests; if values do not stabilize with more paths and reasonable basis refinements; if regression matrices are rank-deficient; if the duality gap remains wide; or if omitted state variables can be shown to alter future conditional payoff.

Alternatives and when they are better

Binomial or trinomial trees are transparent and powerful in low-dimensional Markov settings.

Finite-difference/PDE methods can be highly accurate for one or two continuous state variables.

Primal-dual simulation gives explicit upper and lower bounds and is valuable for verification.

Dynamic programming with richer regression or machine learning can approximate continuation in higher dimensions but introduces new regularization and explainability issues.

Nested Monte Carlo can estimate continuation more directly but is computationally expensive because each current state launches many inner future simulations.

How this connects to the surrounding knowledge estate

LSMC starts with the Monte Carlo path engine and adds an approximate dynamic-programming layer. The low-dimensional comparison owner is the American-option binomial tree. A stochastic-volatility implementation can consume parameters from Heston calibration. The resulting price can be challenged under the broader model-validation framework.

Verification and update triggers

Preserve model parameters, exercise grid, basis specification, state definition, path count, random-number scheme, discounting convention, training/valuation split and regression diagnostics. Revalidate after product-payoff changes, model recalibration, new exercise features, basis-library changes, simulation-engine upgrades, persistent duality gaps or material divergence from independent lattice/PDE benchmarks.

Primary and high-quality references

• Francis A. Longstaff and Eduardo S. Schwartz, Valuing American Options by Simulation: A Simple Least-Squares Approach, Review of Financial Studies, 2001.
• John N. Tsitsiklis and Benjamin Van Roy, Regression Methods for Pricing Complex American-Style Options, IEEE Transactions on Neural Networks, 2001.
• L. C. G. Rogers, Monte Carlo Valuation of American Options, Mathematical Finance, 2002.
• Leif Andersen and Mark Broadie, Primal-Dual Simulation Algorithm for Pricing Multidimensional American Options, Management Science, 2004.
• Chenru Liu, Jaehyuk Choi and others, Leave-One-Out Least Squares Monte Carlo Algorithm for Pricing Bermudan Options, Journal of Futures Markets, 2024.

Educational boundary: This article explains numerical optimal stopping and derivative pricing. It does not recommend exercising, buying or selling any option.