A hybrid transformer framework integrating sentiment and dynamic market structure for stock price movement forecasting

Traditional financial forecasting models are inherently limited as their sole reliance on price data causes them to overlook critical information such as market sentiment and dynamic inter-market interactions. To address this shortcoming, this study proposes a novel hybrid transformer model that integrates heterogeneous data sources. Specifically, our framework combines investor sentiment extracted from news headlines using FinBERT and dynamic changes in market structure analyzed with a TVP-VAR model, along with traditional log-return data. In forecasting experiments conducted on four major global stock indices (S&P 500, FTSE 100, CSI 300, and Nikkei 225), the proposed model consistently demonstrated superior performance compared to both single-data-source models and traditional benchmarks. We found that the inclusion of dynamic market structure information, derived from the TVP-VAR model, as a predictive variable was a decisive factor in improving forecasting accuracy. This research empirically validates that a multi-modal approach combining heterogeneous data can significantly enhance the precision of financial market forecasting.