A Dynamic Regime-Switching Model Using Gated Recurrent Straight-Through Units

Abstract

We introduce a novel approach for regime identification using deep learning, a recurrent neural network architecture termed the gated recurrent straight-through unit (GRSTU). The new model can be implemented using commonly available open-source machine learning libraries, enabling automatic differentiation, and trained with the Adam optimizer. Through comprehensive simulation studies, we illustrate that the GRSTU model surpasses statistical jump models, which have shown state-of-the-art performance in regime identification tasks. Specifically, the GRSTU excels in regime classification, particularly on smaller datasets, while demonstrating comparable performance on larger datasets. Finally, in an out-of-sample application, we employ the GRSTU to identify regime changes in the S&P500 index from January 1, 2003 through January 1, 2024. We find that simple regime-switching strategies outperform the index, in terms of lower volatility, CVaR, and drawdown, while maintaining a Sharpe ratio equivalent to or better than that of the baseline.

Keywords: Deep learning, Gated recurrent units, Regime switching, Recurrent neural networks, Statistical jump models, Straight-through estimators, Time series forecasting, Unsupervised learning

JEL Classification: C22, C32, C38, C45, C51, C53, C58, C61, G10, G11, G12, G14