Can Machine Learning Catch Economic Recessions Using Economic and Market Sentiments?
Quantitative models are an important decision-making factor for policy makers and investors. Predicting an economic recession with high accuracy and reliability would be very beneficial for the society. This paper assesses machine learning technics to predict economic recessions in United States using market sentiment and economic indicators (seventy-five explanatory variables) from Jan 1986 - June 2022 on a monthly basis frequency. In order to solve the issue of missing time-series data points, Autoregressive Integrated Moving Average (ARIMA) method used to backcast explanatory variables. Analysis started with reduction in high dimensional dataset to only most important characters using Boruta algorithm, correlation matrix and solving multicollinearity issue. Afterwards, built various cross-validated models, both probability regression methods and machine learning technics, to predict recession binary outcome. The methods considered are Probit, Logit, Elastic Net, Random Forest, Gradient Boosting, and Neural Network. Lastly, discussed different models performance based on confusion matrix, accuracy and F1 score with potential reasons for their weakness and robustness.