Abstract

Deep neural networks for financial time-series forecasting achieve competitive accuracy but are computationally intensive, raising practical and environmental concerns when used at scale in trading and risk management systems. Existing pruning methods compress such models using largely task-agnostic criteria (e.g. weight magnitude), ignoring the specific requirements of financial applications, where behaviour in high-volatility regimes and risk-adjusted performance are critical. This work proposes a post-training structured pruning framework for deep time-series models used in finance. After training, we compute an importance score for each parameter group (channels, attention heads, or neurons) that combines gradient-based sensitivity with a regime-weighted loss, giving higher weight to periods of market stress than to calm periods. Groups with low importance are removed using dependency-aware structured pruning. We implement the framework on several state-of-the-art architectures available in TSLib (e.g. linear, mixer, convolutional, and Transformer-based models) and evaluate them on multiple equity index, stock panel and cryptocurrency datasets. The study aims to provide statistically sound, domain-aware guidelines for building greener financial forecasting systems.