Abstract
Importance weighting is a general way to adjust Monte Carlo integration toaccount for draws from the wrong distribution, but the resulting estimate canbe highly variable when the importance ratios have a heavy right tail. Thisroutinely occurs when there are aspects of the target distribution that are notwell captured by the approximating distribution, in which case more stableestimates can be obtained by modifying extreme importance ratios. We present anew method for stabilizing importance weights using a generalized Paretodistribution fit to the upper tail of the distribution of the simulatedimportance ratios. The method, which empirically performs better than existingmethods for stabilizing importance sampling estimates, includes stabilizedeffective sample size estimates, Monte Carlo error estimates, and convergencediagnostics. The presented Pareto $\hat{k}$ finite sample convergence ratediagnostic is useful for any Monte Carlo estimator.