Transition variable selection for regime switching recurrent reinforcement learning

Non-linear time series models, such as regime-switching (RS), have become increasingly popular in economics. In the literature, regime-switching recurrent reinforcement learning (RS-RRL), a combined technique of statistical modeling and machine learning, has been proposed to build financial trading platforms and enhance trading profits by modeling the nonlinear dynamics of stock returns with smooth transition autoregressive (STAR) models. In this paper, we address the transition variable selection issue in the RS-RRL trading system. Four indicators, namely volume, relative strength index, implied volatility and conditional volatility are considered as possible options for transition variable selection in RS-RRL. Of the four indicators, it is found that the RS-RRL trading system with the volume indicator produces a better Sharpe ratio than others.