Exponential Tilting and Esscher Transforms

Simulating a portfolio's chance of a catastrophic loss can be painfully slow with plain Monte Carlo: if the event you care about only happens once in ten thousand simulated paths, you need an enormous number of runs before you see enough examples to estimate its probability reliably. Exponential tilting is a trick for making the rare event happen far more often in the simulation — while keeping careful track of exactly how much you distorted things, so you can correct back to the true probability afterward.

The idea

Take a distribution's density $f(x)$ and reweight it using an exponential factor:

$$f_\theta(x) = \frac{e^{\theta x} f(x)}{M(\theta)} ,$$

where $M(\theta)$ is the moment generating function (a normalizing constant that makes $f_\theta$ integrate to 1). In plain English: multiplying by $e^{\theta x}$ pushes probability mass toward larger values of $x$ when $\theta$ is positive, shifting the whole distribution's center of gravity toward the tail you care about, while $M(\theta)$ just rescales everything back so it's still a valid probability distribution. Conveniently, tilting a distribution from a well-known family (normal, exponential, Poisson) this way keeps it in the same family, just with shifted parameters — an Esscher-transformed normal is still normal, only with a different mean.

To use this for simulation: instead of drawing from the original distribution and waiting for a rare event to show up naturally, you draw from the tilted distribution $f_\theta$, where that event is common — then you weight each simulated result by $f(x)/f_\theta(x)$ (the "importance weight," which corrects for the fact that you sampled from the wrong distribution on purpose) before averaging. Done correctly, this gives an unbiased estimate of the true rare-event probability, computed from far fewer total simulation paths than plain Monte Carlo would need to see the event happen naturally even a handful of times.

Where this shows up

This is the standard tool behind efficient tail-risk simulation in credit portfolio models, where the event of interest (many defaults happening together) can be extremely rare under the true distribution but needs to be estimated precisely for a capital or risk-limit calculation. Rather than running an infeasible number of naive simulations, tilting the default-generating distribution toward the "many defaults" region and correcting with importance weights gets a precise estimate from a tractable number of simulated paths.