Bank systemic risk prediction based on text mining and explainable machine learning

Abstract

This study utilizes textual data from The Wall Street Journal, employing 12 machine learning models to forecast systemic risk in the US banking sector. Then, this paper applies the SHAP method to interpret the prediction results. The empirical conclusions are as follows: Firstly, in terms of time series forecasting, deep learning models exhibit the best performance, tree models demonstrate moderate predictive efficacy, while linear models perform poorly in predictions. Secondly, there is a positive correlation between SHAP values and banking systemic risk, this conclusion fills the previous research gap. Further research reveals that Topic_29 consistently ranks at the top in feature importance across various time windows. Its keywords (interest rate, bank, stock, company, inflation, rate cut, China) suggest that interest rate policies, corporate operations, inflation control, and geoeconomic factors play pivotal roles in systemic risk. Additionally, the study observes a negative correlation between news sentiment and SHAP values; negative sentiment has a stronger impact and a longer duration. Finally, this study links the topic keywords back to the original news texts to elucidate the impact of news on systemic risk across different sliding window periods.