Monte Carlo Methods and Variance Reduction
Explain how the Monte Carlo method works for estimating expectations. Then describe and compare the main variance reduction techniques: antithetic variates, control variates, importance sampling, stratified sampling, and quasi-Monte Carlo.
For each technique, explain the intuition for why it reduces variance and give a concrete example of when you would use it in practice.
Hints
- Start with the basic MC estimator $\hat{\theta} = (1/N)\sum f(X_i)$ and explain why error scales as $O(1/\sqrt{N})$ via CLT.
- For each variance reduction technique, identify what known structure it exploits: symmetry (antithetic), correlation (control variates), tail importance (IS), spatial coverage (stratification, QMC).
- Give a concrete finance example for each: antithetic pairs for vanilla options, a European call as a control variate for an Asian call, IS for deep OTM option pricing.
Worked Solution
How to Think About It: Monte Carlo is the brute-force workhorse of quantitative finance. Whenever you cannot compute $E[f(X)]$ analytically -- which is most of the time for path-dependent derivatives, high-dimensional integrals, or complex portfolio simulations -- you simulate it. The method is dead simple: draw $N$ samples, compute $f$ for each, average the results. The catch is that the error decays as $O(1/\sqrt{N})$, meaning you need 100x more samples to get 10x more accuracy. Variance reduction is about being smarter with your samples so you get the same accuracy with fewer draws.
Key Insight: All variance reduction techniques exploit some known structure in the problem to reduce the randomness in your estimate. The more you know about $f$ and $X$, the more variance you can squeeze out.
The Method:
1. Basic Monte Carlo
To estimate $\theta = E[f(X)]$, draw $X_1, \ldots, X_N$ i.i.d. from the distribution of $X$ and compute:
$$\hat{\theta} = \frac{1}{N} \sum_{i=1}^N f(X_i)$$
By the law of large numbers, $\hat{\theta} \to \theta$ as $N \to \infty$. By the CLT, the standard error is $\sigma_f / \sqrt{N}$ where $\sigma_f^2 = \text{Var}(f(X))$.
2. Antithetic Variates
Idea: instead of drawing $N$ independent samples, draw $N/2$ pairs $(U_i, 1 - U_i)$ (or more generally, pair each sample with its "mirror image"). If $f$ is monotone, the two evaluations in each pair are negatively correlated, so their average has lower variance than two independent draws.
Estimator: $\hat{\theta} = \frac{1}{N} \sum_{i=1}^{N/2} [f(U_i) + f(1-U_i)]$.
When to use: pricing options where the payoff is monotone in the underlying (e.g., vanilla calls). Pairing an up-path with a down-path is a natural antithetic construction.
3. Control Variates
Idea: find a random variable $g(X)$ whose expectation $E[g(X)]$ you know analytically, and which is correlated with $f(X)$. Subtract off the "error" in $g$ to reduce the error in $f$:
$$\hat{\theta}_{CV} = \hat{\theta} - c(\bar{g} - E[g])$$
where $c = \text{Cov}(f, g) / \text{Var}(g)$ is the optimal coefficient. Variance reduction factor: $1 - \rho^2_{f,g}$.
When to use: pricing an exotic option where a related vanilla option has a known formula. Use the vanilla payoff as the control variate. For example, pricing an Asian call using a European call as control.
4. Importance Sampling
Idea: sample from a different distribution $q(x)$ that puts more weight where $f(x)p(x)$ is large, and correct with likelihood ratios:
$$\hat{\theta}_{IS} = \frac{1}{N} \sum_{i=1}^N f(X_i) \frac{p(X_i)}{q(X_i)}, \quad X_i \sim q$$
The optimal proposal is $q^{*}(x) \propto |f(x)| p(x)$, which gives zero variance (but requires knowing $\theta$, so it is not directly useful). In practice, you choose $q$ to shift mass toward the important region.
When to use: estimating rare-event probabilities (e.g., probability of a 10-sigma move). Standard MC gives almost all zeros -- importance sampling shifts the distribution toward the tail.
5. Stratified Sampling
Idea: partition the sample space into strata $S_1, \ldots, S_K$, then sample proportionally from each stratum. This guarantees you cover the entire space evenly, unlike pure random sampling which can cluster.
$$\hat{\theta}_{SS} = \sum_{k=1}^K P(S_k) \cdot \hat{\theta}_k$$
Variance reduction comes from removing the between-strata variance. The total variance is only the within-strata variance.
When to use: high-dimensional integrals where certain regions contribute disproportionately. For example, stratifying by the first time step in a path simulation.
6. Quasi-Monte Carlo
Idea: replace pseudorandom numbers with low-discrepancy sequences (Sobol, Halton) that fill the space more evenly. The error improves from $O(1/\sqrt{N})$ to $O((\log N)^d / N)$ for $d$-dimensional integrals -- essentially $O(1/N)$ for moderate dimensions.
When to use: pricing high-dimensional derivatives (basket options, mortgage-backed securities). Works best when the integrand is smooth and the effective dimension is moderate.
Practical Considerations:
- Antithetic variates: easiest to implement, modest improvement (typically 20-50% variance reduction for monotone payoffs).
- Control variates: most widely used in practice. The key is finding a good control -- the closer to $|\rho| = 1$, the better.
- Importance sampling: powerful but dangerous. A bad proposal can increase variance dramatically. Requires problem-specific tuning.
- Stratified sampling: useful but harder in high dimensions (curse of dimensionality for the strata).
- Quasi-MC: strong theoretical guarantees, but performance degrades in very high dimensions and with discontinuous integrands.
Answer: Monte Carlo estimates $E[f(X)]$ by sample averaging with $O(1/\sqrt{N})$ convergence. The five main variance reduction techniques each exploit different structure: antithetic variates use symmetry, control variates use correlation with known quantities, importance sampling reweights toward important regions, stratified sampling ensures even coverage, and quasi-MC uses deterministic low-discrepancy sequences.
Intuition
The fundamental bottleneck with Monte Carlo is the $1/\sqrt{N}$ convergence rate -- to halve the error, you need four times the samples. All variance reduction techniques work by injecting known information about the problem to reduce the portion of randomness that is "wasted" on things you already know. Antithetic variates exploit the symmetry of the distribution. Control variates subtract out the noise you can predict analytically. Importance sampling concentrates your computational budget on the part of the space that matters. The common thread: the more prior knowledge you encode, the less you need to brute-force.
In practice on a trading desk, control variates are the most common because there is almost always a simpler derivative with a known price that is correlated with the exotic you are pricing. Importance sampling is the most powerful but also the most dangerous -- a poorly chosen proposal distribution can make things much worse. The skill is knowing which technique to reach for given the problem structure.