Financial Sentiment Analysis for Cross-Market Forecasting, Trading, and Explainability: A Survey and Research Roadmap
Abstract
Text is a first-class signal in equity markets, and the methods used to extract it have moved through three distinct eras: domain specific lexicons, fine-tuned transformer encoders, and general-purpose large language models. In parallel, four downstream literatures earnings call tone, interpretable time series forecasting, cost-aware trading strategy backtesting, and explainable AI in finance have each matured largely in isolation from one another. This survey reviews all four strands together, tabulates the leading methods against one another across nine comparative tables and six figures, and visualises the resulting coverage as an explicit gap map. From those gaps it justifies, point by point, a four-part research program multi-channel sentiment extraction, sentiment fused forecasting, trading strategy construction and cost aware backtesting, channel and role-attributed explainability rather than any of the visible alternative designs. Two contributions distinguish it from existing reviews: a covariate-compatibility analysis showing that the newest state-of-the-art forecasting backbones are architecturally unable to accept a historical sentiment covariate at all, and a two-axis attribution taxonomy separating feature-level, channel-level, and speaker-role-level explanation. The survey is intended as the literature foundation for that four-part program.