Machine Learning Based Portfolio Analysis

Abstract

This paper investigates the effectiveness of machine learning methods in predicting stock returns in China. In this paper, we use 16 liquidity, momentum, market risk and macro fundamentals indicators, as well as 112 interaction terms between indicators and macro variables with monthly stock excess returns for the lagged period from January 2000 to November 2015 as the sample set, train elasticity networks, gradient boosting, random forests, neural networks, and simple linear regression models, and based on the data of December 2015, predict the January 2016 returns, form portfolios based on the predictions and hold them for 12 months, and thereafter roll over the extended sample period to form new portfolios, and evaluate the performance of the portfolios based on out-of-sample alpha versus the estimated shap value to assess the significance of the metrics. The results show that the machine learning approach significantly outperforms the simple linear regression model and performs better in terms of weighted portfolio returns. The portfolio constructed based on the neural network model (NN1) achieves the lowest mean square error and the highest model fit, and NN3 achieves the highest out-of-sample alpha, suggesting that the model is able to achieve higher risk-adjusted returns when risk factors are taken into account. Further analysis shows that the NN3 model is able to effectively identify high-yielding stocks and control portfolio risk, making it a preferred strategy for investment portfolios in the Chinese equity market. This paper still supports Leippold's (2022) view that liquidity indicators are most important after excluding fundamental indicators.