An experimental study on forecasting using TES processes

Forecasting is of prime importance for accuracy in decision making. For data sets containing high autocorrelations, failure to account for temporal dependence will result in poor forecasting. TES (Transform-Expand-Sample) is a class of stochastic processes to model empirical autocorrelated time series and is used in Monte Carlo simulation. Its merit is to simultaneously capture both the empirical distribution function and the autocorrelation function. The transition structure of TES processes can be utilized to calculate forecasts for future periods. In this paper, we utilize phase-type random variables as the innovation density in TES model fitting methodology, and we investigate the forecasting performance of TES processes compared to traditional auto regressive integrated moving-average models. We find that TES models yield forecasts as accurate as time series models.