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

TL;DR

This paper evaluates neural network models like LSTM and GRU for predicting US recessions from 1967 to 2021, demonstrating their effectiveness especially for longer-term forecasts and interpretability through SHAP explanations.

Contribution

It introduces the application of LSTM and GRU models for recession prediction and uses SHAP for interpretability, highlighting their ability to capture business cycle asymmetries and nonlinearities.

Findings

Neural networks outperform traditional models in long-term recession forecasts.

SHAP explanations identify key indicators like S&P 500 and term spread.

Models are robust against other interpretability methods.

Abstract

Long short-term memory (LSTM) and gated recurrent unit (GRU) are used to model US recessions from 1967 to 2021. Their predictive performances are compared to those of the traditional linear models. The out-of-sample performance suggests the application of LSTM and GRU in recession forecasting, especially for longer-term forecasts. The Shapley additive explanations (SHAP) method is applied to both groups of models. The SHAP-based different weight assignments imply the capability of these types of neural networks to capture the business cycle asymmetries and nonlinearities. The SHAP method delivers key recession indicators, such as the S&P 500 index for short-term forecasting up to 3 months and the term spread for longer-term forecasting up to 12 months. These findings are robust against other interpretation methods, such as the local interpretable model-agnostic explanations (LIME) and the marginal effects.