Improving futures market predictions with sentiment using a pre-trained robustly-optimized bidirectional encoder representations from transformers approach in China

Abstract

This study bridges operational research and financial analysis to investigate the interdependencies between price discovery processes and media sentiments related to commodities in the Chinese futures market. Across a data-rich span from January 2018 to May 2023, we systematically aggregate and assess daily sentiment data extracted from a diverse array of media sources. By adapting the Robustly-optimized Bidirectional Encoder Representations from Transformers pretraining approach, our Operational Research processing improves traditional sentiment scoring methods, enabling a holistic assimilation of semantic data that encapsulates both influential and consequential market factors. Our empirical investigation focusing on the predictive modelling of Chinese commodity futures’ returns, identifies sentiment as a key extra element, also confirming the most relevant factors already identified in the literature, showing an enhanced adjustment effect of sentiment in the Chinese market compared to findings in other international futures markets. The incorporation of extensive semantic analysis into our Operational Research models affords a fresh vantage point upon sentiment’s role in shaping market behaviours, thus informing some past conclusions. This confluence of Operational Research textual analysis methodology and financial analysis enriches the available toolkit for market strategists, and offers a new conduit for academic and practitioner inquiries into financial markets’ dynamics.