Can transformers transform financial forecasting?

Purpose This study aims to critically evaluate the competitiveness of Transformer-based models in financial forecasting, specifically in the context of stock realized volatility forecasting. It seeks to challenge and extend upon the assertions of Zeng et al. (2023) regarding the purported limitations of these models in handling temporal information in financial time series. Design/methodology/approach Employing a robust methodological framework, the study systematically compares a range of Transformer models, including first-generation and advanced iterations like Informer, Autoformer, and PatchTST, against benchmark models (HAR, NBEATSx, NHITS, and TimesNet). The evaluation encompasses 80 different stocks, four error metrics, four statistical tests, and three robustness tests designed to reflect diverse market conditions and data availability scenarios. Findings The research uncovers that while first-generation Transformer models, like TFT, underperform in financial forecasting, second-generation models like Informer, Autoformer, and PatchTST demonstrate remarkable efficacy, especially in scenarios characterized by limited historical data and market volatility. The study also highlights the nuanced performance of these models across different forecasting horizons and error metrics, showcasing their potential as robust tools in financial forecasting, which contradicts the findings of Zeng et al. (2023) Originality/value This paper contributes to the financial forecasting literature by providing a comprehensive analysis of the applicability of Transformer-based models in this domain. It offers new insights into the capabilities of these models, especially their adaptability to different market conditions and forecasting requirements, challenging the existing skepticism created by Zeng et al. (2023) about their utility in financial forecasting.