Trading Strategies Based on Machine Learning Predictions of Long-Term Stock Price Movement in the Healthcare Industry

Abstract

While a variety of current research involves the use of machine learning algorithms to predict stock price movement, the existing research largely uses stock indices to create short-term predictions for trading purposes. I seek to explore whether machine learning algorithms can accurately make long-term predictions for stocks in specifically the healthcare industry, and whether these predictions can effectively be applied to trading strategies that ultimately help the portfolio beat healthcare indices. I apply industry-specific features in order to predict stock price, testing multiple datasets by varying the features used and varying the prediction time windows. Using the predictions based on the optimal features and time window, I then implement the market cap weighted, equal weighted, inverse volatility weighted, minimum variance weighted, and factor model weighted strategies, varying them by testing different rebalancing periods. I evaluate the combined trading strategies and prediction methods by comparing the absolute return and Sharpe Ratio to that of the S&P 500 Index (SPX) and the iShares Dow Jones U.S. Healthcare ETF (IYH). I found that while SVM is a machine learning method that outperforms other methods in terms of prediction accuracy, SVM may not necessarily also outperform other methods when applied to a trading strategy. Rather, my results suggest that the Lasso and Decision Tree models actually outperform the other models when applied to a trading strategy. This discrepancy may be due to the use of the models for classification, rather than for predicting a numerical magnitude of stock price movement.