and a policy imperative. For years, financial economists have relied on autoregressive (AR) time series models to predict market movements, due to their simplicity, transparency, and generally reliable performance. However, traditional AR models often fall short when it comes to capturing the complexity of modern financial systems, particularly during extreme events where nonlinearities and feedback loops dominate. This raises an important question: can artificial intelligence (AI) and machine learning (ML) models outperform these classical approaches in financial forecasting?
A recent working paper from the Bank for International Settlements (BIS) provides evidence in this respect. The paper, “Predicting Financial Market Stress with Machine Learning,” introduces market condition indicators (MCIs) for three critical U.S. markets—Treasury, foreign exchange (FX), and money markets—and uses these to compare the forecasting performance of classical time series models with tree-based ML models, particularly Random Forests. The results are striking: across multiple horizons, ML models substantially outperform AR models in predicting the upper quantiles of financial stress. For instance, in FX markets, Random Forests reduce quantile loss by as much as 27% compared to AR models when forecasting the 90th percentile of stress levels at a three-month horizon. Similar improvements are observed in money markets, especially beyond the two-month horizon.
The key advantage of machine learning lies in its ability to model complex, nonlinear relationships and interactions between variables. Traditional AR models operate under linear assumptions and tend to overfit when extended into multivariate frameworks. In contrast, ML models handle high-dimensional data and sparse signals more effectively. This makes them particularly valuable for forecasting tail risks—those rare but highly impactful stress events that policymakers and investors are most concerned about.
One of the most persistent and legitimate criticisms of machine learning (ML) models—especially in high-stakes domains like finance—is their lack of interpretability. This refers to the difficulty in understanding how these models arrive at their predictions. Unlike traditional statistical models such as linear regressions, where the relationship between inputs (independent variables) and outputs (dependent variables) is explicit and governed by clear equations, ML models—particularly ensemble methods like random forests or deep learning architectures—are often seen as “black boxes.” They can learn highly complex, non-linear relationships, but at the cost of transparency. For policymakers, regulators, and economists, this is problematic because decision-making in these fields requires not just accurate forecasts, but also explanations that can be linked to economic intuition and real-world mechanisms.
To address this challenge, the BIS researchers use a technique called Shapley values, which originates from cooperative game theory, a field that studies how to fairly distribute gains (or costs) among a group of players working together. In the context of ML, the “players” are the input features (i.e., predictors), and the “gain” is the model’s prediction. The Shapley value provides a way to equitably attribute the contribution of each feature to a particular prediction, taking into account all possible combinations of features and how they interact.
In trying to predict stress in the FX market , a Random Forest model might use dozens of variables—ranging from interest rate spreads to fund flows and volatility indicators. The Shapley value method assigns each of these variables a numerical score that reflects how much it contributed, positively or negatively, to the model’s prediction in that specific case. This isn’t just based on correlation or variable importance in a global sense; it reflects the local impact of each variable on each prediction.
What makes Shapley values powerful is that they respect three key fairness properties:
1. Efficiency – The total predicted value is distributed across all features.
2. Symmetry – If two features contribute equally in all contexts, they receive equal credit.
3. Dummy – If a feature has no marginal contribution, it gets a zero value.
By using Shapley values, the BIS researchers bring transparency to their ML model. They can explain which factors drive the forecasted market stress, and by how much. For instance, in forecasting financial stress:
• Investor overextension, captured through variables like mutual fund flows into high-yield assets or equity funds, consistently ranks high in Shapley value importance. This suggests that when investors allocate aggressively into riskier segments—often during calm periods—they may be sowing the seeds of future stress.
• The global financial cycle (e.g., synchronized movements in global capital flows and risk appetite) also shows up as a major predictor. This supports the idea that financial vulnerabilities are not just domestic phenomena—they are globally interconnected.
• Importantly, the previous values of the Market Condition Indicators (MCIs) themselves are highly predictive of future stress. This reflects what the authors call self-reinforcing dynamics—a condition where stress in a market today increases the likelihood of stress tomorrow. It could be due to mechanisms like liquidity spirals, where initial stress reduces market liquidity, which then amplifies the next shock, and so on.
Moreover, the study uncovers strong cross-market linkages, or what they term stress spillovers. For example, elevated stress in Treasury markets can predict future stress in FX markets, and vice versa. This is not surprising, given the interconnectedness of modern financial systems—many institutions operate across markets, and disruptions in one segment (like a spike in Treasury yields) can force them to rebalance, liquidate, or hedge in others (like the FX market).
In summary, the BIS researchers’ use of Shapley values bridges the gap between machine learning’s predictive power and the interpretability needs of policy and finance. It allows for actionable insights: identifying not just when stress is likely to occur, but also why, and where to look. This makes machine learning not only a forecasting tool but also a diagnostic instrument for systemic risk surveillance.
While AI and ML models show clear advantages in capturing complexity and improving forecast accuracy, especially for tail events, their effectiveness can depend on the specific application and data structure. Hybrid approaches that integrate the robustness and interpretability of classical models with the predictive power of machine learning seem to hold the most promise. The BIS study and others suggest that we are entering a new era of financial forecasting—one where machine learning is not just a supplement to traditional methods, but in many cases, a superior replacement.
In conclusion, AI and ML models shows great promise in outperforming autoregressive time series models in forecasting financial market stress. Their ability to detect subtle patterns, accommodate nonlinearities, and adapt to rapidly changing conditions makes them indispensable tools for the next generation of financial analysts, risk managers, and central bankers. The findings of this study are crucial for financial stability monitoring. They imply that not only should central banks and regulators look at current stress levels, but also at predictive indicators of build-up, particularly in the tail ends of the risk distribution. And thanks to Shapley values, they can now trace those signals back to specific drivers—like investor behavior or market structure changes—rather than simply relying on opaque machine learning outputs.
References
Aldasoro, Iñaki, Peter Hördahl, Andreas Schrimpf, and Xingyu Sonya Zhu. 2025. Predicting Financial Market Stress with Machine Learning. BIS Working Paper No. 1250. Bank for International Settlements.
Leave a Reply