Volatility and value-at-risk forecasting using BERT and transformer models incorporating investors' textual sentiments

Effective risk management contributes to the stability of financial markets. This study investigates how decomposed investor sentiments affect financial volatility, emphasizing the role of sentiment polarity. Using BERT to extract positive and negative sentiment indices from stock forums, we find that negative sentiment has a stronger effect on realized volatility than positive sentiment, reflecting asymmetric investor reactions. Integrating these decomposed sentiment indices with technical indicators into a Transformer model significantly improves the forecasting of both realized volatility and value-at-risk (VaR). Our findings underscore the economic relevance of distinguishing sentiment polarity and offer novel insights into combining fine-grained sentiment features with deep learning for financial risk forecasting.