Employing Machine Learning Algorithms to build Trading Strategies with higher than Risk-Free Returns

This research aims at exploring whether simple trading strategies developed using state-ofthe-art Machine Learning (ML) algorithms can guarantee more than the risk-free rate of return or not. For this purpose, the direction of S&P 500 Index returns on every 6th day (SPYRETDIR6) and magnitude of S&P 500 Index daily returns (SPYMAG) were predicted on a broad selection of independent variables using various ML techniques. Using five consecutive data spans of equal length, GBM was found to provide highest prediction accuracy on SPYRETDIR6, consistently. In terms of magnitude prediction of daily returns (SPYMAG), Random Forest results indicated that there is a very high correlation between actual/predicted values of SPY. Based on these results, Trading Strategy #1 (using SPYRETDIR6 predictions) and Trading Strategy #2 (using SPYMAG predictions) were developed and tested against a simple Buy & Hold benchmark of the same index. It was found that Trading Strategy #1 provides negative returns on all data spans, while Trading Strategy #2 has positive returns on average when data is separated into consecutive data spans. None of the trading strategies have a positive Sharpe ratio on average, but Trading Strategy #2 is almost as profitable as investing in T-bills using the risk-free rate.

Keywords: Machine Learning, S&P 500, Forecasting, Ensemble Methods, XGBoost.
JEL Classifications: C32, C45, C52, C53, C55.
DOI #: 10.33818/ier.805042