Volatility-Sensitive Forecasting : Deep Learning Model Robustness across Calm and Crisis Periods

Vancsura, László and Tatay, Tibor and Bareith, Tibor (2026) Volatility-Sensitive Forecasting : Deep Learning Model Robustness across Calm and Crisis Periods. Virtual Economics, 9 (1). pp. 28-64. ISSN 2657-4047

Preview

Text Vancsuraetal_2026_9_1_2.pdf - Published Version Available under License Creative Commons Attribution. Download (1MB) | Preview

Abstract

This study investigates how the forecasting performance of deep learning models is affected by changing market conditions, with particular emphasis on periods of differing volatility. While deep learning methods are increasingly applied in financial forecasting, relatively limited evidence exists on how their predictive accuracy varies across distinct market regimes. The aim of this paper is to examine whether model performance systematically deteriorates during more turbulent periods and whether this effect differs across model architectures. The empirical analysis covers multiple asset classes, including equities, commodities, cryptocurrencies, and currency pairs, over three partially overlapping periods (2016–2018, 2018–2020, and 2020–2022), capturing both relatively stable and highly volatile market environments. Forecasting performance is evaluated using several recurrent neural network architectures, and the relationship between market volatility and prediction error is analysed using cross-sectional regression techniques. Additional sensitivity analyses are conducted to assess the robustness of the results across alternative asset samples. The results reveal a consistent positive relationship between market volatility and forecasting error, indicating that predictive performance tends to deteriorate under more turbulent market conditions. However, the magnitude of this effect varies across model architectures. In particular, GRU-based architectures appear less sensitive to volatility, whereas the use of multivariate inputs does not consistently improve forecasting accuracy. These findings suggest that greater model complexity does not necessarily translate into more robust performance under changing market conditions. The study is subject to several limitations, including the relatively small cross-sectional sample and the use of standardised model configurations. Future research could extend the analysis by considering alternative model classes, broader datasets, and more comprehensive robustness frameworks.

Actions (login required)

Edit Item