Volatility-aware market-based indicators for weekly systemic-risk forecasting: evidence from U.S. financials

Bošnjak, Mile, Pejić Bach, Mirjana
Central European Journal of Operations Research · Springer Nature · 2026-11-20 · DOI: 10.1007/s10100-026-01054-6

Аннотация

This paper develops and evaluates a market-based framework for forecasting weekly systemic risk in the U.S. financial sector. Using daily equity prices for a sample of major U.S. financial institutions since 2012, we construct a composite Systemic Risk Index (SRI) that combines realised volatility, market co-movement, and market-wide risk aversion as measured by the VIX. One-week-ahead forecasts of the SRI are generated using ElasticNet, Extreme Gradient Boosting (XGBoost), Artificial Neural Networks (ANN), and a benchmark autoregressive specification within an expanding-window walk-forward evaluation framework. The predictor set incorporates realised variance, rolling correlations, maximum drawdowns, EGARCH(1,1) conditional volatility estimates, and lagged values of the SRI. Forecasting performance is assessed using both the Area Under the Receiver Operating Characteristic Curve (AUC) and the Root Mean Squared Error (RMSE), together with sensitivity, ablation, robustness, and statistical-significance analyses. The results show that market-based indicators possess substantial predictive content for future systemic stress. ElasticNet consistently achieves the strongest overall performance, closely followed by XGBoost, while more complex specifications do not generate systematic forecasting gains. Additional fuzzy-volatility transformations provide little incremental predictive value and fail to improve performance significantly. The main findings remain stable across alternative stress thresholds, training-window lengths, and model-update frequencies. From an operational research perspective, the proposed framework can be viewed as a transparent decision-support system for financial-stability monitoring. The study therefore contributes to the operational-research literature on early-warning systems by demonstrating how high-frequency market information can be transformed into an interpretable and reproducible financial-stability monitoring tool. The results suggest that parsimonious and interpretable forecasting models may offer a more effective balance between predictive accuracy, robustness, and operational feasibility than increasingly complex machine-learning architectures.

Читать у издателя →

Похожие статьи