e-Archivo. Repositorio Institucional de la Universidad Carlos III de Madrid
Publicación
Artículo preliminar (WorkingPaper)
Information, Not Flexibility: Attributing Forecast Gains in Neural Stochastic Volatility
When a neural volatility model beats a benchmark, is it the network or the data? We change one ingredient at a time within one Bayesian stochastic volatility framework, for three stock indices at one-, five-, and ten-day horizons. It is the data. Adding realised variance lowers the quasi-likelihood (QLIKE) loss by 2.8% to 13.4% in all nine cases and improves the one-day log score in every market. A neural network does not beat a linear function in one-day QLIKE with identical inputs, even inputs of the heterogeneous autoregressive (HAR) type. Combining the two neural models adds no stable gain. HAR and Realized GARCH have the lowest QLIKE in the nine-model comparison. With Student-t innovations, linear return feedback makes multi-step variance means infinite, while a bounded network does not. In practice, add realised information and choose the innovation distribution carefully, because it changes one-day tail risk rankings. Flexibility comes last., J.M. Marín acknowledges financial support from the Agencia Estatal de
Investigación through project PID2020-115598RB-I00 and from Universidad
Carlos III de Madrid through project 2024/00701/001. Helena Veiga acknowledges financial support from the Agencia Estatal de Investigación through projects PID2025-168576NB-I00 and PID2022-139614NB-C22, from the Fundação para a Ciência e a Tecnologia through project UIDB/00315/2020, and from Universidad Carlos III de Madrid through project 2024/00701/001.