Monte Carlo Pricing: Buying Precision Without Fooling Yourself

Monte Carlo simulation estimates a financial quantity by averaging outcomes across modelled scenarios. It is flexible enough for multiple risk factors and complex path dependence, but flexibility does not remove the need to specify the measure, contractual timing and error budget.

This reference combines basic valuation simulation with variance reduction, correlated paths and dynamic-hedging experiments. The examples are retained or explicitly constructed teaching calculations, not realised client performance or current market predictions.

Start with the quantity being estimated

For discounted payoff $Y$, a simulation estimate based on $N$ independent paths is

$$\hat V_N=\frac1N\sum_{i=1}^N Y_i.$$

Under appropriate integrability, the average converges to the model expectation. If the payoff has finite variance, estimated standard error is approximately $s_Y/\sqrt N$, where $s_Y$ is the sample standard deviation of discounted payoffs.

The probability measure matters. Pricing simulations use an appropriate pricing measure and discounting convention. A physical-measure risk simulation answers a different question about possible outcomes. Using a historical expected return in a basic risk-neutral option-pricing simulation creates a different calculation even if the code is otherwise identical.

A point estimate should be accompanied by its statistical uncertainty. The retained numerical example uses estimate 12.5, payoff standard deviation 4 and 10,000 independent observations. Standard error is 0.04, giving the usual approximate 95% interval of 12.42 to 12.58. This interval reflects sampling error under the model, not uncertainty about whether the model is correct.

A vanilla benchmark exposes implementation errors

For a non-dividend-paying asset under basic risk-neutral geometric Brownian motion,

$$S_T=S_0\exp[(r-\tfrac12\sigma^2)T+\sigma\sqrt T Z],$$

where $Z$ is standard normal, $r$ is the constant discount rate and $\sigma$ is annual volatility. The discounted call payoff is $Y=e^{-rT}\max(S_T-K,0)$, with strike $K$.

This simple case has an analytical benchmark, so it is a useful first check. Simulated terminal mean should approach $S_0e^{rT}$, and the option estimate should approach the matching analytical value within its standard error.

If only terminal price matters, generating a full daily path is unnecessary. It increases cost without improving the exact terminal distribution. A path is needed when the payoff depends on intermediate observations or when several time-dependent state variables must be evolved together.

Path dependence creates a second approximation

The observed prices on the retained arithmetic-average path

Their average is 102.5, so an average-price call with strike 100 pays 2.5 before discounting. These values illustrate one path; pricing requires averaging such discounted payoffs over many paths.

The monitoring schedule is contractual. A daily average and a continuous-time average are different payoffs. A continuously monitored barrier can be crossed between simulation dates, so endpoint checks can miss knockout events. Increasing path count reduces sampling error but leaves that monitoring bias.

Brownian-bridge corrections, refined monitoring or specialised sampling can address missed crossings under suitable models. Exact transitions remove state-transition bias at sampled dates, but they do not automatically recover every event between those dates.

Correlation belongs in the scenario generator

For a desired correlation matrix $R$, factorise $R=LL^\top$ and transform independent standard normals $Z$ into $LZ$. The transformed vector has covariance $R$. A positive-definite matrix supports a standard Cholesky factorisation; a semidefinite matrix may require a suitable eigenvalue-based treatment.

The matrix must be symmetric with unit diagonal and valid eigenvalues. Pairwise correlation estimates can fail to form a valid joint matrix because of missing observations or inconsistent estimation windows. Repairing that matrix is a modelling intervention that should be measured and documented.

Correlated Brownian innovations do not imply that terminal arithmetic returns have exactly the same correlation under every nonlinear model. The dependence of model drivers and the dependence of transformed outputs should be distinguished.

For a basket payoff $\max(\sum_i w_iS_i(T)-K,0)$, correlation changes the distribution of the weighted sum. For spread and worst-of payoffs, its effect can differ. Correlation sensitivity should be calculated for the actual payoff rather than inferred from a universal slogan about diversification.

Variance reduction targets the estimator

Antithetic sampling pairs a draw $Z$ with $-Z$ and averages the two payoffs. It helps when the paired outputs have negative covariance. It is not guaranteed to help every payoff, and evaluating the second payoff has a computational cost. The relevant comparison is error per unit of total computation.

A control variate uses a related variable $C$ with known expectation $\mu_C$:

$$\hat V_{\mathrm{CV}}=\bar Y+\lambda(\mu_C-\bar C).$$

The population-optimal coefficient is $\lambda=\operatorname{Cov}(Y,C)/\operatorname{Var}(C)$. A highly correlated analytical vanilla value or geometric-average payoff can be useful when estimating a more complex but related payoff. The control’s expected value must actually be known under the same model.

Stratification divides the sampling space into regions and allocates samples deliberately. Importance sampling instead draws from a proposal density $q$ and weights by the likelihood ratio $p/q$, where $p$ is the target density. The proposal must cover the relevant support, and unstable weights can destroy the expected benefit.

Quasi-Monte Carlo uses more evenly distributed points. Randomised versions can support uncertainty estimates through independent randomisations. The ordinary independent-sample standard-error formula should not be applied blindly to a deterministic low-discrepancy point set.

The square-root convergence cost is material

Ordinary Monte Carlo standard error falls in proportion to $N^{-1/2}$. The retained example shows that reducing the error of 25,000 paths by half requires approximately 100,000 paths, all else equal. Doubling paths reduces standard error by about 29%, rather than by 50%.

Parallel computation speeds up the same sampling task, but it does not change this statistical convergence rate. Random-number streams must remain independent or appropriately managed across workers. Reusing an identical seed on every worker can create repeated paths masquerading as a larger sample.

Reproducibility records should include generator settings, model parameters, path count, grid, payoff version and discounting conventions. A seed is useful for reproducing a run, but it is not evidence that a particular run is representative.

Dynamic-hedging simulations require a cash account

A delta-hedging experiment must account for the option position, underlying holdings, cash financing and transaction costs. Rebalancing the underlying without updating the cash account creates or destroys wealth artificially.

The simulation should record option gamma and underlying level at each hedge date because local variance effects are weighted by the exposure present when a move occurs. Final realised volatility alone cannot reconstruct the hedge ledger. The broader economic interpretation is developed in The Volatility Risk Premium: Insurance Pricing, Market Regimes and the Cost of Convexity.

The historical teaching discussion compared hedging with implied volatility and with an estimate of actual volatility under strong assumptions. Preserve that comparison as two explicitly specified simulation policies. Each policy must use only information available at its hedge date, and both must face the same paths, financing and costs. Future realised volatility cannot be supplied retrospectively as an available input.

A retained historical chart compares three-month at-the-money implied volatility with a 60-observation historical-volatility measure for a US large-cap equity index over 2015–2019, using an annualisation factor of 260. It illustrates that the two series can diverge and react at different speeds. Their different horizons and constructions prevent the visible gap from being interpreted directly as an executable profit. The chart’s commercial interface and identifying labels are omitted; its methodological distinction is retained.

For each policy, retain the initial option cash flow, underlying trades, cash-account interest, final settlement and transaction costs. Reconcile their sum to the reported profit or loss. This makes the experiment reproducible and separates a difference in hedging policy from a difference in accounting conventions.

Greeks and tail risk need their own precision tests

Finite-difference sensitivities can reuse the same random numbers in bumped and unbumped valuations to reduce noise in the difference. This common-random-number technique is distinct from increasing path count. Bump size balances truncation error against simulation noise.

Pathwise and likelihood-ratio methods offer alternative sensitivities when their regularity assumptions hold. Discontinuous payoffs and exercise decisions may complicate differentiation. A low-noise price estimate does not guarantee a low-noise gamma estimate.

Tail quantiles can converge slowly because only a small fraction of paths inform the tail. Expected shortfall additionally depends on the severity of those losses. Report effective tail observations and sensitivity to distributional assumptions rather than presenting extra decimal places as extra knowledge.

The final error record should separate sampling uncertainty, state discretisation, monitoring error, regression or exercise approximation and model uncertainty. Monte Carlo is reliable when each component is identified; a large path count addresses only one of them.

Early exercise adds an estimation problem inside simulation

For an American-style option, a path’s realised future payoff is not information available at an earlier exercise date. Exercising whenever that realised future payoff would be lower creates look-ahead and overstates the holder’s opportunity.

A regression-based method instead estimates continuation value from state variables available at the exercise date. The exercise rule compares immediate payoff with that estimated conditional value. Basis functions, training paths and exercise dates become additional modelling choices.

Using separate paths to evaluate the fitted exercise policy helps distinguish policy quality from in-sample regression optimism. A low regression error does not automatically imply a small pricing error near the exercise boundary, where small errors can change the selected action. High-dimensional states and sparse exercise regions can make the continuation estimate especially difficult.

Importance-sampling weights require direct inspection

If a proposal distribution oversamples a rare default or barrier event, the likelihood ratio must restore the original probability weighting. Omitting it changes the target expectation. A proposal can also be poorly chosen even when the formula is correct: a few very large weights can dominate the estimate.

Inspect weight concentration, variance and sensitivity to proposal parameters. Repeating the simulation with independent seeds can reveal instability, but repeated failures to sample an influential region can still create false confidence. A useful proposal covers all outcomes that materially contribute to the target.

The objective is lower estimator variance per unit of computation. It is not to make the simulated event happen more often for its own sake. A proposal that increases event frequency while producing extreme weights may provide little or no improvement.

A reproducible sensitivity experiment

Suppose a price sensitivity is estimated by central difference, $(V(x+h)-V(x-h))/(2h)$, for input $x$ and bump $h$. Independent path sets in the two prices can make the difference much noisier than either price. Common random numbers align the scenarios and often cause much of that noise to cancel.

The bump cannot be reduced indefinitely. A tiny bump magnifies residual sampling and numerical error after division by $2h$; a large bump measures a broad nonlinear change rather than a local derivative. Running several economically sensible bump sizes reveals whether the sensitivity is stable.

If changing the input requires rebuilding a calibrated curve or volatility surface, the sensitivity definition should state whether that rebuild is included. A frozen-parameter sensitivity and a recalibrated market-quote sensitivity answer different questions even when both are called delta or vega.

Report the model question alongside the estimate

A simulation report should state what is averaged, under which probabilities, with which discounting and contractual dates. Its numerical interval should be labelled as sampling uncertainty. Parameter ranges, alternative dependence assumptions and stressed paths belong alongside that interval when they materially affect the decision.

This presentation prevents a narrow confidence interval from being mistaken for comprehensive certainty. Simulation can precisely estimate an expectation inside a weak model. Its strength is the ability to make assumptions and scenarios explicit; its precision is valuable only when those assumptions match the intended financial question.