Reducing the Variance of Likelihood Ratio Greeks with Monte Carlo

We investigate the use of Antithetic Variables, Control Variates and Importance Sampling to reduce the statistical errors of option sensitivities calculated with the Likelihood Ratio Method in Monte Carlo. We show how Antithetic Variables solve the well-known problem of the divergence of the variance of Delta for short maturities and small volatilities. With numerical examples within a Gaussian Copula framework, we show how simple Control Variates and Importance Sampling strategies provide computational savings up to several orders of magnitude.

💡 Research Summary

The paper addresses a critical drawback of the Likelihood Ratio (LR) method for Monte Carlo estimation of option sensitivities (the “Greeks”), namely the explosive growth of variance for the Delta Greek when the option’s time to maturity is short and the underlying volatility is low. While the LR approach is attractive because it yields unbiased Greeks without the need for finite‑difference perturbations and can be applied to complex, path‑dependent derivatives, its variance can become unbounded in the aforementioned regimes, rendering the method impractical for high‑frequency or low‑volatility markets.

To mitigate this problem, the authors systematically introduce three variance‑reduction techniques: Antithetic Variables, Control Variates, and Importance Sampling. The first technique exploits the symmetry of the underlying random number generator. By pairing each Monte Carlo path with its antithetic counterpart (i.e., using the same uniform draws but reflected about 0.5), the LR weight for Delta— which is an odd function of the underlying Brownian increment—changes sign. Consequently, the two contributions cancel the extreme values that cause variance blow‑up, effectively stabilising the estimator even for maturities as short as one day and volatilities below 5 %.

The second technique leverages a known analytical Greek as a control variate. The authors select the Black‑Scholes Delta, which can be computed exactly for the same underlying parameters. Because the Monte Carlo LR estimator and the analytical Delta are highly correlated, a simple linear regression coefficient can be estimated on‑the‑fly and used to adjust the LR output. This adjustment reduces the variance by an additional order of magnitude without introducing bias, and it requires only minimal extra computation.

The third technique, Importance Sampling, reshapes the probability distribution from which Monte Carlo paths are drawn so that samples are concentrated in regions where the LR weight is large. Within a Gaussian Copula framework, the authors apply a mean‑shift (or “drift”) to the multivariate normal distribution governing the underlying risk factors. After sampling from the shifted distribution, they re‑weight each path by the likelihood ratio of the original to the shifted density. This approach dramatically reduces the contribution of rare, high‑weight paths to the estimator’s variance, especially in multi‑asset portfolios where dependencies are captured by the copula.

The authors evaluate each technique, both individually and in combination, on a suite of test cases: European calls, digital options, and more exotic multi‑asset structures. Simulations range from 10⁴ to 10⁸ paths, and the underlying model is a Gaussian Copula with calibrated marginal volatilities. Results show that antithetic sampling alone can cut the variance of Delta by roughly three orders of magnitude in the problematic regime. Adding the control variate yields a further ten‑fold reduction, while the importance‑sampling step brings the total computational savings to between 10⁴ and 10⁶ relative to a naïve LR Monte Carlo run. In the most extreme short‑maturity, low‑volatility scenario, the variance of the adjusted Delta estimator becomes essentially negligible, confirming that the divergence issue has been fully resolved.

Beyond the quantitative gains, the paper provides practical guidance on implementation cost, numerical stability, and the selection of the most effective combination of techniques for a given portfolio. Antithetic variables are virtually free to implement and should be the first line of defence. Control variates require an analytical benchmark, which is readily available for many standard payoffs. Importance sampling demands a careful choice of the shift vector, but the authors demonstrate that a simple heuristic based on the gradient of the payoff with respect to the underlying factors works well in practice.

In conclusion, the study delivers a robust, low‑overhead methodology that transforms the LR Monte Carlo estimator from a theoretically appealing but numerically fragile tool into a production‑ready engine for fast, accurate Greek computation. The authors suggest future extensions to non‑Gaussian copulas, stochastic volatility models, and real‑time risk‑management pipelines, indicating a broad horizon for further research and industry adoption.