Predicting Stock Market Movements Using Machine Learning Techniques

The purpose of this paper is to compare the performance of various state-of-the-art machine learning techniques in predicting the behavior of stock-market returns. To do so, we gathered ten years of daily historical data (2488 observations per stock) for the top ten most liquid stocks in Casablanca Stock Exchange (Morocco) and trained six machines learning classifiers (ridge regression, LASSO regression, support-vector machine, k-nearest neighbors, random forest, and adaptive boosting) and an ensemble of them (i.e. ensemble learning) in order to predict one-day-ahead, one-week-ahead, and one-month-ahead prices direction (i.e. positive or negative returns). The performance of each algorithm is then evaluated using accuracy, precision, recall, and F1 scores. Applying the Diebold-Mariano test at a significance level of 5%, we have found that support-vector machine, random forest, and adaptive boosting perform equally well and outperform all other single classifiers for short-term predictions (one-day-ahead and one-week-ahead). However, for monthly predictions, all methods display similar predictive accuracy. In addition, our study suggests that ensemble learning significantly improves all performance metrics for the three prediction horizons. We have also found that for all models the performance significantly decreases as the prediction horizon increases.The purpose of this paper is to compare the performance of various state-of-the-art machine learning techniques in predicting the behavior of stock-market returns. To do so, we gathered ten years of daily historical data (2488 observations per stock) for the top ten most liquid stocks in Casablanca Stock Exchange (Morocco) and trained six machines learning classifiers (ridge regression, LASSO regression, support-vector machine, k-nearest neighbors, random forest, and adaptive boosting) and an ensemble of them (i.e. ensemble learning) in order to predict one-day-ahead, one-week-ahead, and one-month-ahead prices direction (i.e. positive or negative returns). The performance of each algorithm is then evaluated using accuracy, precision, recall, and F1 scores. Applying the Diebold-Mariano test at a significance level of 5%, we have found that support-vector machine, random forest, and adaptive boosting perform equally well and outperform all other single classifiers for short-term predictions (one-day-ahead and one-week-ahead). However, for monthly predictions, all methods display similar predictive accuracy. In addition, our study suggests that ensemble learning significantly improves all performance metrics for the three prediction horizons. We have also found that for all models the performance significantly decreases as the prediction horizon increases