Predicting macro-financial instability – How relevant is sentiment? Evidence from long short-term memory networks

Abstract

This paper examines the relevance of sentiment in predicting overall financial system instability using long-run short-term memory networks. Weekly data on the US financial system, consumer sentiment, producer sentiment, and investor sentiment is collected from 21 January 1994 to 27 December 2019, and different models are developed to predict the one-week-ahead levels of financial stress in the US financial system. We find that models using sentiment indices outperform those relying solely on historical financial stress and risk data. This result is robust to comparisons with an alternative deep learning method and out-of-sample predictions. It constitutes an argument in favor of behavioral finance and Minsky’s (Knell, 2015) financial instability hypothesis against the Efficient Market Hypothesis. As it concretely identifies the main indicators for predicting US financial stress one week in advance, the study provides relevant recommendations for policymakers and investors in terms of macroprudential policies and portfolio management.