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 this paper, researchers from the Bank for International Settlements explore the potential of machine learning (ML) to provide more accurate and timely predictions of financial market stress, focusing on key US markets.
They 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. They 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, they identify key factors that contribute to future market stress, such as funding liquidity, investor overextension and global financial cycles.
Researchers found that 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 behavior 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, the findings demonstrate the potential of machine learning in financial forecasting, particularly in complex and dynamic environments.