Inside the black box: Neural network-based real-time prediction of US recessions

Résumé 0

A standard feedforward neural network (FFN) and two specific types of recurrent neural networks, long short-term memory (LSTM) and gated recurrent unit (GRU), are used for modeling US recessions in the period from 1967 to 2021. The estimated models are then employed to conduct real-time predictions of the Great Recession and the Covid-19 recession in the US. Their predictive performances are compared to those of the traditional linear models, the standard logit model and the ridge logit model. The out-of-sample performance suggests the application of LSTM and GRU in the area of recession forecasting, especially for the long-term forecasting tasks. They outperform other types of models across five different forecast horizons with respect to a selected set of statistical metrics. Shapley additive explanations (SHAP) method is applied to GRU and the ridge logit model as the best performer in the neural network and linear model group, respectively, to gain insight into the variable importance. The evaluation of variable importance differs between GRU and the ridge logit model, as reflected in their unequal variable orders determined by the SHAP values. These different weight assignments can be attributed to GRUs flexibility and capability to capture the business cycle asymmetries and nonlinearities. The SHAP method delivers some key recession indicators. For forecasting up to 3 months, the stock price index, real GDP, and private residential fixed investment show great short-term predictability, while for longer-term forecasting up to 12 months, the term spread and the producer price index have strong explanatory power for recessions. These findings are robust against other interpretation methods such as the local interpretable model-agnostic explanations (LIME) for GRU and the marginal effects for the ridge logit model.