Decision-Aware Conditional GANs for Time Series Data

Abstract

We introduce the decision-aware time-series conditional generative adversarial network (DAT-CGAN), a method for the generation of time-series data that is designed to support decision-making. The framework adopts a multi-Wasserstein loss on decision-related quantities and an overlapped block-sampling approach for sample efficiency. We characterize the generalization properties of DAT-CGAN and in application to a multi-period portfolio choice problem and financial time series data, we demonstrate better training stability and generative quality in regard to both raw data and decision-related quantities than strong GAN-based baselines.