Predicting financial market stress with machine learning

Summary

Focus

Understanding and predicting financial market stress is crucial for maintaining economic stability. Episodes of market stress can disrupt credit availability, influence asset prices and hinder economic growth. Traditional measures of financial conditions, such as financial stress indices (FSIs) and financial conditions indices (FCIs), often fail to distinguish between general market sentiment and specific vulnerabilities, reducing their predictive power. In our paper, we explore the potential of machine learning to provide more accurate and timely predictions of financial market stress, focusing on key US markets.

Contribution

We develop new market condition indicators (MCIs) for three key US markets: Treasury, foreign exchange (FX) and money markets. These indicators focus on market-specific issues like liquidity problems and deviations from standard no-arbitrage conditions. We use tree-based machine learning models, specifically random forests, to predict the future distribution of these MCIs. These models are shown to be more accurate than traditional time series methods, especially for predicting extreme stress scenarios (the "tails" of the distribution). Additionally, using Shapley value analysis, we identify key factors that contribute to future market stress, such as funding liquidity, investor overextension and global financial cycles.

Findings

Random forests significantly outperform traditional methods in predicting financial market stress, especially over longer horizons (3–12 months). Key predictors of market stress include factors related to liquidity, investor behaviour and the global financial cycle. The new MCIs provide valuable real-time insights into market-specific stress that traditional indices might miss. For policymakers, these tools offer powerful means of monitoring and predicting financial market stress and guiding targeted interventions. For researchers, our findings demonstrate the potential of machine learning in financial forecasting, particularly in complex and dynamic environments.

Abstract

Using newly constructed market condition indicators (MCIs) for three pivotal US markets (Treasury, foreign exchange, and money markets), we demonstrate that tree-based machine learning (ML) models significantly outperform traditional timeseries approaches in predicting the full distribution of future market stress. Through quantile regression, we show that random forests achieve up to 27% lower quantile loss than autoregressive benchmarks, particularly at longer horizons (3–12 months). Shapley value analysis reveals that funding liquidity, investor overextension and the global financial cycle are important predictors of future tail realizations of market conditions. The MCIs themselves play a prominent role as well, both in the same market (self-reinforcing dynamics within markets) and across markets (spillovers across markets). These results highlight the value of ML in forecasting tail risks and identifying systemic vulnerabilities in real time, bridging the gap between highfrequency data and macroeconomic stability frameworks.

JEL classification: G01, C53, G17, G12, G28

Keywords: machine learning, financial stress, quantile regressions, forecasting, Shapley value