Evidence-grounded AI risk event representation for listed-company credit risk early warning: a multi-source RAG and reliability-gated fusion approach

AI-related corporate narratives may describe substantively different events, including operational adoption, product commercialization, investment expansion, speculative promotion, implementation failure, and regulatory concern. These heterogeneous narratives may therefore have different predictive associations with future corporate credit risk. Existing text-enhanced models mainly rely on sentiment, keywords, or document-level embeddings, which may fail to distinguish substantive AI adoption from speculative disclosure, implementation failure, or regulatory concern. This study proposes an evidence-grounded AI risk event framework for listed-company credit risk prediction. The framework constructs firm-year evidence libraries from AI-related news, corporate announcements, annual-report management discussion and analysis, investor-interaction texts, financial indicators, and macroeconomic variables. A retrieval-augmented generation (RAG) module retrieves firm-specific evidence, and a large language model (LLM) converts retrieved evidence into structured AI risk events. A temporal reliability-gated fusion network subsequently integrates financial fundamentals, macroeconomic conditions, event-level AI semantics, source availability, extraction confidence, and cross-source consistency. Using 31,320 firm-year observations of Chinese listed companies from 2008 to 2017, the proposed model achieves an RMSE of 4.665 natural distance-to-default units, an MAE of 2.919, an \(\:{R}^{2}\) of 0.423, a standard-deviation-normalized RMSE of 0.760, and a high-risk AUC of 0.791 under walk-forward validation. The incremental improvement over the strongest component-level comparator using the same RAG event vectors and an FT-Transformer is smaller than the improvement produced by event-level representation itself: the paired firm-block bootstrap yields a mean squared-loss difference of 0.81 with a 95% confidence interval of [0.03, 1.68] and \(\:p=0.044\) . Predictive discrimination is also heterogeneous with respect to evidence availability, with AUCs of 0.817 and 0.748 for held-out observations with and without a valid AI event, respectively. An independent evaluation of 240 firm-year evidence libraries containing 2,684 manually identified relevant passages yields a Precision@8 of 0.842, Recall@8 of 0.781, nDCG@8 of 0.861, and firm-year event recall of 0.836. Ablation, robustness, and parameter-sensitivity tests confirm the value of evidence retrieval, event-level representation, source consistency, temporal attention, and reliability gating. Later-period, alternative-taxonomy, source-removal, temporal-knowledge, and firm-block bootstrap analyses further support the stability of the findings. The results indicate that traceable event-level representations of AI-related texts provide incremental predictive information in the studied historical sample. Most of the observed gain relative to structured and conventional text representations is attributable to evidence-grounded event construction, whereas reliability-conditioned temporal fusion provides a smaller additional improvement. These findings represent predictive associations rather than causal effects of corporate AI narratives on credit risk.