Monte Carlo Methods for Quant Interviews: Simulation, Variance Reduction, Pricing

Monte Carlo Methods for Quant Interviews: Simulation, Variance Reduction, and Practical Pricing

Monte Carlo methods are everywhere in quant finance. Pricing exotic options, computing risk metrics (VaR, expected shortfall), backtesting strategies, simulating portfolios, and validating models — all rely on simulation when closed-form solutions are unavailable or too restrictive. For quant-research and quant-developer interviews, Monte Carlo questions probe whether candidates understand simulation as a tool, not just a button to press, and whether they can reason about variance, convergence, and computational efficiency.

This guide covers the core Monte Carlo concepts that show up in interviews, plus the variance-reduction techniques that distinguish candidates with practical experience from those who’ve only seen textbook examples.

The Core Idea

To estimate E[f(X)] where X is a random variable with known distribution, draw N independent samples X_1, …, X_N and compute the sample mean (1/N) Σ f(X_i). By the law of large numbers, this converges to the true expectation as N → ∞.

Standard error: σ_estimator = σ(f(X)) / √N. Convergence is slow (1/√N): to halve the standard error you need 4× the samples. This is the central tension in Monte Carlo — it’s flexible but slow. Most practical work involves variance-reduction techniques to extract more precision per sample.

Pricing an Option by Monte Carlo

Canonical example: price a European call option on a stock following geometric Brownian motion under the risk-neutral measure.

Algorithm:

1. Draw N samples of standard normal: Z_1, …, Z_N.
2. Compute terminal stock prices: S_T(i) = S_0 exp((r – σ²/2)T + σ√T · Z_i).
3. Compute payoffs: V_i = max(S_T(i) – K, 0).
4. Discount: estimate price = (1/N) Σ exp(-rT) V_i.

For a vanilla European call, this is overkill (Black-Scholes has a closed form). For path-dependent options (Asian, barrier, lookback) or basket options, Monte Carlo is essential.

Standard error estimate: compute sample standard deviation of (exp(-rT) V_i) and divide by √N.

Variance Reduction Techniques

Antithetic variates

For each draw Z_i, also use -Z_i in the simulation. Pair the results: V_i’ = (V(Z_i) + V(-Z_i)) / 2. If V is monotonic in Z, this reduces variance because the two paired draws are negatively correlated.

Practical impact: for a European call, antithetic variates reduces standard error by roughly a factor of 2 with no additional sampling cost beyond the negation.

Control variates

Use a related quantity Y with known mean E[Y] to reduce variance. Define V_corrected = V – β(Y – E[Y]) where β is chosen optimally to minimize variance. Optimal β is Cov(V, Y) / Var(Y).

Example: pricing an Asian option. Use the geometric average (which has a closed form) as a control variate for the arithmetic average (which doesn’t). The two are highly correlated; the closed-form version pins down the offset.

Importance sampling

If the function f(X) is concentrated where the original distribution puts little mass (rare events), sampling from the original distribution wastes most samples. Sample instead from a distribution that puts more mass where f matters; correct using likelihood-ratio weighting.

Example: pricing a deep out-of-the-money option, where most paths give zero payoff. Shift the drift to make the option more likely to expire in the money; correct via Radon-Nikodym derivative. This dramatically reduces variance for tail-event simulations.

Stratified sampling

Divide the sample space into strata; ensure each stratum is sampled proportionally. For one-dimensional problems, Latin hypercube sampling generalizes this to higher dimensions.

Quasi-random / low-discrepancy sequences

Replace pseudo-random numbers with deterministic low-discrepancy sequences (Sobol, Halton, Niederreiter). Convergence improves to roughly O(1/N) instead of O(1/√N) under suitable conditions. Particularly effective for moderate-dimensional integrals (typically up to ~30 dimensions).

Common Interview Questions

Estimate π by Monte Carlo

“How would you estimate π using Monte Carlo?” Generate N points uniformly in [0,1]². Count the fraction within the unit quarter-circle. Multiply by 4. This is the textbook example; the interviewer is probably warming up before deeper questions.

Price a barrier option

“Price a knock-out call by Monte Carlo.” Simulate paths step by step (multiple time steps required for path dependence). Check at each step whether the barrier is breached; if so, set payoff to zero. Average discounted payoffs. Discuss bias from discrete checking vs continuous barriers (Brownian bridge correction).

Compute Greeks by Monte Carlo

“How do you estimate delta (∂V/∂S) by Monte Carlo?” Three approaches: (a) finite differences (re-simulate at S+ε and S-ε; compute slope; biased and noisy); (b) likelihood ratio method (works for general payoffs but high variance); (c) pathwise method (works for smooth payoffs; differentiable through the path). Strong candidates know multiple approaches and trade-offs.

Discuss variance reduction

“You’re pricing an option with Monte Carlo and the standard error is too high. What do you do?” Increase sample size (slow). Apply antithetic variates (cheap, often effective). Apply control variates (more thought required, often very effective). Consider importance sampling for tail events. Move to quasi-random sequences for moderate dimensions. Strong candidates list multiple options and discuss when each is appropriate.

Discuss VaR estimation

“How do you estimate 99% VaR by Monte Carlo?” Simulate portfolio P&L under risk factors; take the 1st percentile of the loss distribution. Discuss issues: tail estimation requires many samples; rare events require importance sampling; correlation structure between risk factors must be modeled correctly.

Address random number generation

“What random number generator would you use?” Modern: Mersenne Twister, PCG, or xoshiro. For cryptographic applications, use a CSPRNG. For financial Monte Carlo, period and statistical properties matter more than cryptographic security. Don’t use rand() (poor quality). Discuss seed management for reproducibility.

Practical Considerations

Convergence diagnostics

Don’t just report a single MC estimate. Compute standard error. Run multiple replicates with different seeds and check stability. For path-dependent simulations, increase the number of time steps and check that the answer stabilizes.

Computational efficiency

Vectorize where possible (NumPy / Eigen). Parallelize across samples (embarrassingly parallel for most MC). Watch for memory: long-horizon, high-frequency simulations can blow up RAM. Profile your simulation; the inner loop usually dominates.

Reproducibility

Always seed the RNG explicitly. For parallel simulations, use independent streams (e.g., counter-based RNGs like Philox) to avoid correlations between threads.

Bias vs variance

Path-dependent options simulated with finite time steps have discretization bias plus statistical variance. Both matter. Strong candidates discuss both: increase time steps to reduce bias; increase sample count to reduce variance.

Frequently Asked Questions

See also: Stochastic Calculus for Quant Interviews • Options Pricing for Quant Interviews • Regression and Hypothesis Testing for Quant Interviews