AI is now setting the gold standard for banking risk analysis: a sweeping 15-year study of 134 banks in 16 European countries finds that neural networks outperform legacy models in identifying systemic threats. Published in Journal of Risk and Financial Management under the title “Forecasting Systemic Risk in the European Banking Industry: A Machine Learning Approach”, the research evaluates the comparative effectiveness of artificial intelligence models, including artificial neural networks (ANNs) and support vector machines (SVMs), alongside traditional econometric models such as AR-GARCH.
The study leverages advanced statistical metrics to assess how well these models predict systemic risk, defined as the risk of collapse in the entire banking system due to the failure of one or more interconnected institutions. It uses two established risk indicators, Delta Conditional Value at Risk (∆CoVaR) and Marginal Expected Shortfall (MES), to evaluate both the contribution of individual banks to systemic risk and their exposure to it.
How is systemic risk measured in the European banking sector?
To quantify systemic risk, the researchers use ∆CoVaR to estimate how much a bank contributes to system-wide risk during financial distress, and MES to measure how much a bank stands to lose during system-wide downturns. These metrics were calculated weekly for each of the 134 banks using stock return data from Bloomberg.
The dataset includes commercial, diversified, and investment banks across major EU economies, including Germany, France, Italy, and the United Kingdom. According to the study, systemic risk levels peaked during periods of financial instability, such as the 2008 global financial crisis and the 2010–2011 European sovereign debt crisis. Notably, Greece and Ireland exhibited the highest average levels of both systemic risk contribution and exposure.
By measuring these risk indicators, the study not only tracks historical fluctuations in systemic vulnerability but also establishes a foundation for testing the predictive power of various forecasting models.
Can machine learning accurately predict future risk?
The research then turns to predictive modeling, comparing three approaches: ANNs, SVMs, and the widely-used AR-GARCH model. Using weekly ∆CoVaR and MES values from 2002 to 2015 as input data, the researchers forecast systemic risk levels for each bank across 2016. The actual observed values from 2016 serve as a benchmark to evaluate model accuracy.
Two ANN architectures were tested, one with a single hidden layer and one with two. The deeper, two-layer ANN outperformed all other methods in forecasting both systemic risk contribution and exposure. For ∆CoVaR, the two-layer ANN yielded an adjusted mean absolute percentage error (A-MAPE) of 11.5%, while the SVM achieved 10.26%. However, for MES, the two-layer ANN again proved strongest with a lower A-MAPE than both SVM and AR-GARCH.
The results clearly indicate that deeper ANN structures are more capable of capturing the nonlinear, complex, and interdependent characteristics of systemic risk in financial systems. In contrast, SVMs exhibited inconsistent performance, and GARCH models, though widely used in volatility forecasting, underperformed significantly, particularly in capturing the contagion dynamics and cross-institutional dependencies that drive systemic instability.
What do these findings mean for financial stability and policy?
The implications are substantial for regulators, policymakers, and financial institutions. The ability to forecast systemic risk with reasonable accuracy opens a path to earlier interventions, better macroprudential regulation, and more targeted capital buffer policies.
The authors emphasize that systemic risk is deeply nonlinear and interlinked, driven not only by volatility but by behavioral responses, contagion effects, and market sentiment. Traditional models like AR-GARCH, which treat each institution in isolation and assume consistent volatility dynamics, are insufficient for capturing the multi-dimensional nature of systemic crises. The study thus supports a pivot toward AI-based models in regulatory toolkits.
From a macroprudential policy standpoint, the findings advocate for the integration of advanced AI-driven early warning systems into existing regulatory frameworks. By monitoring the evolving risk profile of banks in near real time, central banks and financial authorities could implement preemptive measures, ranging from capital requirement adjustments to liquidity injections, before risk cascades into a full-blown crisis.
In addition, the study’s methodology, based on publicly available stock return data and macroeconomic indicators, demonstrates that real-time systemic risk monitoring is both feasible and scalable. The authors note that while SVMs have promise in certain classification tasks, their performance degrades in unstable environments where market conditions shift rapidly.