Equity portfolio optimization using reinforcement learning: An emerging market case

Abstract

In this study, RL models are generated to address the equity portfolio optimization problem. The asset universe for this problem has been restricted to BIST30 constituents, which is a stock market index from an emerging market. The index is composed of the 30 largest stocks in terms of market capitalization. The characteristics of an emerging market do differ from developed markets as emerging markets bring new challenge such as increased uncertainty and decreased market efficiency. Our models use return, volatility, technical indicators, and fundamental company information to make portfolio weight decision. We study DDPG and PPO methods in this study. Two different neural network architectures are designed to fit the needs of these two RL models. We utilized convolutional network structures to successfully extract meaningful information out of the dataset. We train the models with 10 years of daily data, and then test with 1 years of data. We set main benchmark as the market capitalization weighted BIST30 index, as we try to compete it and provide a better weighting strategy. We also compared our agents with UBAH, UCRP and MPT portfolios as additional benchmarks. The results show that we outperform all of these benchmarks in terms of return, and risk adjusted return metrics, such as Sharpe and Sortino ratios. One aspect of the models is that they produce a higher volatility behavior, however they compensate the high risk with a higher return value per taken risk so the return based metrics are superior to benchmarks. We highlight the model-free nature of the our proposed RL agents. The results are presented at the last section and a future projection is provided for a probably better model.