Analytical solutions of the performance of optical communication systems are difficult to obtain and often, Monte Carlo simulations are used to achieve realistic estimates of the performance of such systems. However, for high performance systems, this technique requires a large number of simulation trials for the estimates to be in a reasonable interval of confidence, with the number of trials increasing linearly with the performance of the system. Various variance reduction schemes for Monte Carlo simulations, known as Importance Sampling have been employed to significantly reduce the number of simulation trials for radio communication systems, but have found little application in the optical domain. We develop an Importance Sampling technique to estimate the performance of direct detection optical systems, where the "gain" of Importance Sampling over Monte Carlo simulations is shown to increase linearly with the system performance.