Risk Management Strategy and Management

Stability-oriented variance forecasting for financial panels under market-state shifts

Sözen, Çağlar, Sözen, Mervenur
Springer Nature · 2026-10-29 · DOI 10.1057/s41283-026-00249-0

Abstract

Financial variance forecas/ts can perform differently across loss functions and market states, so pooled accuracy alone may not fully characterize forecasting reliability. We develop and evaluate a stability-oriented predictor-screening framework for one-day-ahead Rogers–Satchell variance forecasting across 23 exchange-traded funds (ETFs). The empirical evidence shows that model rankings depend materially on the evaluation criterion. A heterogeneous autoregressive (HAR) benchmark estimated by ordinary least squares delivers the lowest pooled root mean squared error (RMSE) and mean absolute error (MAE) and the strongest blockwise RMSE ranking. An economically screened random forest achieves the lowest quasi-likelihood (QLIKE) loss, while the stability-screened random forest ranks second under QLIKE and yields significantly lower QLIKE than an unrestricted random forest. However, its squared-error improvement over the unrestricted random forest is not significant after multiple-testing adjustment. Stress periods increase forecast errors across all models, and parsimonious ridge-type specifications exhibit smaller proportional deterioration from calm to stress conditions. These results indicate that stability-oriented screening can reduce dimensionality and preserve competitive nonlinear forecasting performance, particularly under QLIKE, but simple persistence-based and economically informed models remain essential benchmarks. For practitioners, model choice should therefore depend on the loss function and the form of robustness that matters for the forecasting task.

Read at publisher →