Stock Selection by Machine Learning in Thailand Stock Market (SET)

This research studies and compares the use of different machine learning tools—K-means clustering, Hierarchical clustering, Affinity propagation clustering, and the Random Forest model—with the Low-Correlation Strategy for selecting stocks to create a portfolio on the Stock Exchange of Thailand (SET). Each stock is given equal weight and evaluated by the Sharpe ratio. The data used for clustering and training to predict stock returns includes adjusted closing prices and six financial ratios: Return on Total Assets (ROA), Return on Equity (ROE), Net Profit Margin, Earnings Per Share (EPS), Dividend Yield, and Debt to Equity Ratio (D/E Ratio). This data spans from January 1, 2020, to December 31, 2022, for the in-sample period, and the adjusted closing stock price data from January 1, 2023, to December 31, 2023, for the in-sample forecasts. During the in-sample period, it was found that using Machine Learning to select stocks, both Unsupervised Learning and Supervised Learning created portfolios with higher Sharpe ratios than the Low-Correlation Strategy. On the other hand, when evaluating the Sharpe ratio for the in-sample forecast period using 2023 data, the Sharpe ratio of portfolios using Unsupervised Learning was negative for all portfolios. The Sharpe ratio of the Random Forest portfolio get rapidly decreased. In contrast, the Sharpe ratio of the Low-Correlation Strategy portfolio increased to 0.61. In conclusion, using Machine Learning, both Supervised and Unsupervised Learning, led to higher Sharpe ratios than the Low-Correlation Strategy during the in-sample period. However, in the in-sample forecast period, the portfolios from both types of Machines Learning had lower Sharpe ratios than those from the Low-Correlation Strategy. However, the Random Forest Model of this study still has high errors due to data and economic factors.