A1403
Title: Fitting well but filtering poorly: Residual diagnostics for recurrent neural networks in volatility modeling
Authors: Theo Berger - University of Applied Sciences Hannover (Germany) [presenting]
Abstract: Machine learning models for financial volatility forecasting are predominantly evaluated using prediction error metrics such as RMSE or out-of-sample Value-at-Risk backtests. This evaluation practice is fundamentally insufficient and potentially misleading. Using simulated and empirical financial time series, evidence demonstrates that recurrent neural networks (RNNs) can achieve low prediction errors while simultaneously failing to capture the serial dependence structure of financial returns, a property that is central to any valid volatility model. Specifically, a systematic trade-off emerges: RNN configurations that minimize in-sample RMSE tend to produce residuals that exhibit significant autocorrelation, indicating that the underlying dependence structure remains unfiltered. This trade-off is particularly pronounced when asymmetric volatility dynamics are present, where RNNs achieve their lowest prediction errors precisely in settings where residual diagnostics fail most severely. Standard VaR backtests are further shown to be insufficiently sensitive to detect this misspecification, meaning that a model can pass regulatory scrutiny while being econometrically invalid. The findings call for residual-based diagnostic analysis as a mandatory step in the evaluation of ML models in financial econometrics, analogous to the standards applied to classical GARCH-type models.
