用机器学习预测金融市场压力（英）

Predicting financial marketstress with machinelearning by Iñaki Aldasoro, Peter Hördahl, Andreas Schrimpf andXingyu Sonya Zhu Monetary and Economic Department March 2025 JEL classification: G01, C53, G17, G12, G28. Keywords: Machine learning, financial stress, quantileregressions, forecasting, Shapley value. BISWorking Papers are written by members of the Monetary and EconomicDepartment of the Bank for International Settlements, and from time to time by othereconomists, and are published by the Bank. The papers are on subjects of topicalinterest and are technical in character. The views expressed in them are those of theirauthors and not necessarily the views of the BIS. This publication is available on the BIS website (www.bis.org). ©Bank for International Settlements 2025. All rights reserved. Brief excerpts may bereproduced or translated provided the source is stated. Predicting financial market stress with machine learning∗ I˜naki AldasoroBISPeter H¨ordahlBISAndreas SchrimpfBIS & CEPRXingyu Sonya ZhuBISFebruary, 2025 Abstract Using newly constructed market condition indicators (MCIs) for three pivotal USmarkets (Treasury, foreign exchange, and money markets), we demonstrate thattree-based machine learning (ML) models significantly outperform traditional time-series approaches in predicting the full distribution of future market stress. Throughquantile regression, we show that random forests achieve up to 27% lower quantileloss than autoregressive benchmarks, particularly at longer horizons (3–12 months).Shapley value analysis reveals that funding liquidity, investor overextension and theglobal financial cycle are important predictors of future tail realizations of marketconditions. The MCIs themselves play a prominent role as well, both in the samemarket (self-reinforcing dynamicswithin markets) and across markets (spilloversacrossmarkets).These results highlight the value of ML in forecasting tail risksand identifying systemic vulnerabilities in real time, bridging the gap between high-frequency data and macroeconomic stability frameworks. JEL Codes: G01, C53, G17, G12, G28.Keywords: machine learning, financial stress, quantile regressions, forecasting, Shapleyvalue. 1Introduction Financial market stress is a persistent threat to macroeconomic stability, with cascadingeffects on credit provision, asset prices, and economic growth.The Great FinancialCrisis (GFC), the Covid “dash for cash” and recurring episodes of market illiquidityunderscore the systemic risks posed by unstable and malfunctioning financial markets.Such episodes often originate in seemingly isolated segments – such as money markets orFX swaps – before propagating globally, as interconnected intermediaries and leveragedinvestors amplify shocks.1 Policymakers and academics alike have long sought tools to measure and forecastthese stress dynamics in real time.Traditional approaches, including financial stressindices (FSIs) and financial conditions indices (FCIs), provide aggregate snapshots ofmarket health but often conflate broad sentiment shifts (e.g., equity volatility via theVIX) with structural vulnerabilities like liquidity shortages or arbitrage breakdowns.This conflation limits their usefulness in identifying market-specific stress, which is crit-ical for targeted interventions. Addressing these gaps requires a framework that priori-tizes real-time data and accommodates non-linear dynamics – a task uniquely suited tomachine learning (ML). This paper makes two interrelated contributions. First, we construct novel marketcondition indicators (MCIs) for three pivotal US markets: Treasury, foreign exchange(FX) and money markets. Unlike traditional indices, the MCIs emphasize market mi-crostructure dislocations as reflected in episodes of illiquidity and deviations from no-arbitrage conditions that reflect the balance sheet constraints of intermediaries and theimpairment of arbitrage. Second, we employ tree-based artificial intelligence (AI) mod-els (random forest (RF)) to forecast the full distribution of future market conditionsvia quantile regressions (Koenker and Bassett, 1978; Adrian et al., 2019).Our resultsshow that ML models consistently outperform classical time-series approaches (e.g., au-toregressive and multivariate quantile regressions), particularly at longer horizons (3–12months). A well-recognized drawback of AI/ML models is lack of explainability, i.e. thechallenge of understanding how complex models arrive at their output, and which inputvariables play a meaningful role in producing that output.We rely on Shapley valuesto address this issue (Shapley, 1953).This is critical from a policy perspective as itcan inform which variables help explain shifts in the forecast distribution of MCIs. Wefind evidence that investor overextension (e.g., fund flows, the global financial cycle) and intermediary liquidity constraints (e.g., primary dealer security holdings) are importantdrivers of such distributional shif