Probabilistic Multistep Time Series Forecasting Using Conditional Generative Adversarial Networks

Abstract

Time series forecasting is a problem that has been studied for many years due to the impact it can have on the world economy and well-being. Predicting multiple future values is an especially complex problem due to the increasing error. This is why there is a need to design and evaluate more and better methods for this forecasting problem. The adversarial generative networks seem to have an excellent performance generating time series indistinguishable from real series. It has been shown that a probabilistic prediction of time series called ForGAN adversary generative network has been successfully used for one-step-ahead predictions. In this work, a modified architecture of ForGAN with multiple outputs is proposed in order to perform multiple-step-ahead predictions. We show by means of experiments using a real dataset that statistically significant improvement of multiple-step-ahead predictions with the proposed modified architecture of ForGAN compared with the use of the original ForGAN network is achieved, decreasing RMSE by 17.6% and CRPS by 17.3% when predicting 5 steps ahead.