Forecasting Recessions in the US Economy Using Machine Learning Methods

A quantitative analysis of socio-economic characteristics, the set of which is typical in the pre-crisis periods of a market economy, is carried out. An indicator for forecasting the onset of a recession in the US economy over the next 6, 12 and 24 months has been constructed using machine learning methods (k-nearest neighbors, support vector machine, fully connected neural network, LSTM neural network, etc.). Using roll forward cross-validation, it is shown that the smallest error in predicting the onset of future recessions was obtained by a fully connected neural network. It is also shown that all three constructed indicators successfully predict the onset of each of the last six recessions that occurred in the United States from 1976 to 2021 (Early 1980s recession, Recession of 1981–82, Early 1990s recession, .COM bubble recession, Great Recession, COVID-19 recession). The resulting indicators can be used to assess future economic activity in the United States using current macroeconomic indicators.