Bank for International Settlements working paper finds random forest models improve forecasts of US market stress by up to 27%

The Bank for International Settlements published a working paper that develops new, market-specific indicators for stress in three core US markets (Treasury, foreign exchange and money markets) and tests whether machine learning can better forecast the distribution of future stress than standard time-series models. Using quantile forecasting, the paper finds that tree-based models, particularly random forests, outperform traditional approaches in predicting tail risks over medium horizons. The market condition indicators (MCIs) are constructed from daily measures of volatility, liquidity and no-arbitrage dislocations and are aggregated using a rolling-window principal component approach to avoid look-ahead bias. In out-of-sample tests on monthly averages, a multivariate quantile regression with 44 predictors performs better in-sample but worse out-of-sample than a simpler autoregressive benchmark, consistent with overfitting, while random forests achieve materially lower quantile loss in FX and money markets, including up to 27% lower quantile loss for the 90th-quantile FX stress forecast at a three-month horizon. Shapley value analysis points to funding liquidity conditions, investor overextension measures, the global financial cycle and the MCIs themselves (within-market persistence and cross-market spillovers) as key drivers of forecasted tail stress, while gains for Treasury market forecasts are more limited and not consistently significant. The paper states that the MCI series are being made available to researchers alongside the publication.