Machine Learning Models Accurately Predict Stock Market Crashes Using Macroeconomic Indicators

Abstract

Stock market crashes have serious consequences for individuals, businesses, and national economies. The ability to predict such events in advance would be of considerable value to investors and policymakers alike. In this study, machine learning algorithms were used to assess whether monthly macroeconomic indicators are capable of predicting U.S. stock market crashes at a six-month forward horizon. A crash was defined as a decline of 20% or more in the S&P 500 index from its most recent peak, consistent with the conventional definition of a bear market. Monthly data spanning from January 1950 to December 2023 were retrieved from the Federal Reserve Economic Data (FRED) database and other publicly available sources. Ten macroeconomic features were used as inputs to the models, including the yield curve spread, the unemployment rate, the CBOE Volatility Index (VIX), and the Shiller CAPE ratio. Various machine learning algorithms were utilized, including logistic regression, decision trees, random forest, support vector machines (SVM), and a multi-layer perceptron (MLP). All models were optimized using grid search algorithms with cross validation. The random forest classifier was particularly accurate after optimization, achieving an area under the receiver operating characteristic curve (AUC) of 0.88. Feature importance analysis identified the yield curve spread and the VIX as the most predictive features across all models.