Predicting Stock Market Crashes with Interpretable Machine Learning

Can machine learning-generated stock market crash predictions be amenable to human interpretation? In their June 2021 paper entitled “Explainable AI (XAI) Models Applied to Planning in Financial Markets”, Eric Benhamou, Jean-Jacques Ohana, David Saltiel and Beatrice Guez apply a gradient boosting decision tree (GBDT) to 150 technical, fundamental and macroeconomic inputs to generate daily predictions of short-term S&P 500 Index crashes. They define a crash as a 15-day S&P 500 Index return below its historical fifth percentile within the training dataset. The 150 model inputs encompass:

1. Risk aversion metrics such as asset class implied volatilities and credit spreads.
2. Price indicators such as returns, major stock index Sharpe ratios, distance from a long-term moving average and and equity-bond correlations.
3. Financial metrics such as 12-month sales growth and price-to-earnings ratio forecasts.
4. Macroeconomic indicators such Citigroup regional and global economic surprise indexes.
5. Technical indicators such as market breath and index put-call ratio.
6. Interest rates such as 10-year and 2-year U.S. Treasury yields and break-even inflation level.

They first rank and filter the 150 inputs based on GBDT to discard about two thirds of the variables. They then apply the Shapley value solution concept to identify the most important of the remaining variables and thereby support interpretation of methodology outputs. Using daily values of the 150 model inputs and daily S&P 500 Index roll-adjusted futures prices from the beginning of January 2003 through mid-January 2021 (with data up to January 2019 used for training, the next year for validation and the rest for testing), they find that:

(more…)