A framework of deep reinforcement learning for stock evaluation functions

Abstract

Quantitative trading is a crucial aspect of money management; however, conventional trading strategies are based on indicators and signals, despite the fact that position sizing is arguably the most important issue. In this study, we present a stock evaluation function that outputs the size of the stock in each fixed period as well as the consequences of increasing or decreasing the size of one’s position. The difficulties involved in using machine learning to adjust stock weighting can be attributed to difficulties in obtaining definite answers via supervised learning. We therefore train our evaluation function using reinforcement learning via CNN within the EIIE network architecture and have the agent adjust the size of the position with the purpose of maximizing profits. Back testing was performed using the top 50 stocks in Taiwan, based on market capitalization. In experiments, most of the stock returns outperformed conventional strategies in terms of cumulative stock value.