详情
MEASURING INVESTOR SENTIMENT WITH LARGE LANGUAGE MODELS FOR SYSTEMIC RISK EARLY WARNING IN THE FINANCIAL SECTOR: AN EMPIRICAL STUDY BASED ON THE INFORMER MODEL
This paper examines whether institution-level investor sentiment extracted from Chinese stock message boards improves early warning of systemic risk in the banking sector. Using 1.47 million posts for 14 listed banks from 2008-2024, the study instruction-tunes an open-source large language model (Llama-3-Chinese-8B) via LoRA to build a domain-specific sentiment classifier. From model-labeled texts, four institution-day measures are constructed—(i) total attention (post volume), (ii) polarity balance, and polarity weighted by (iii) views and (iv) comments—capturing both direction and intensity of online sentiment.
Systemic risk is proxied by ΔCoVaR and complemented by DY connectedness measures. For forecasting, the paper employs the Informer architecture tailored to long-horizon time-series, integrating macroeconomic variables and bank-specific characteristics with the sentiment features. On held-out sentiment data, the fine-tuned classifier outperforms BERT and Llama-based baselines across accuracy, precision, recall, and F1. In out-of-sample risk prediction, augmenting features with the proposed sentiment measures reduces MAE and MSE for most banks at both 1-step-ahead and 5-step-ahead horizons. Relative to the best non-Informer baseline, Informer attains improvements of approximately 10–12% in MSE and 7–10% in MAE, with Diebold–Mariano tests indicating statistically significant differences.
The findings document that LLM-based institution-level sentiment especially intensity-weighted indicators, contains incremental and forward-looking information for systemic risk monitoring. The paper provides a micro-anchored “AI+finance” pathway for macroprudential early warning and supports the integration of text-driven analytics into supervisory toolkits.