Answers are grounded in this paper and cite their sources.
Purpose This study develops event-conditional forecasts of bond-market responses to rating disclosures, using a large language model (LLM) to structure report information and statistical models to estimate subsequent outcomes. Design/methodology/approach LLM-assisted extraction is combined with rating histories and pre-announcement market conditions. The sample comprises 19,871 linked rating events and 119,683 eligible event–bond observations in China's exchange-traded bond market. A retrospective expanding-window design evaluates forecasts of transaction-price formation, repricing intensity, trading attention, cross-bond propagation and credit direction out of sample. Findings Cross-bond propagation shows the most consistent gains over the market-state benchmark. For 20-day transaction-price formation, the full-information model lowers Brier scores from 0.2148 to 0.1507 in the issuer-expanded channel and from 0.1501 to 0.1077 for directly rated bonds relative to the event-metadata benchmark. The remaining rating variables provide no statistically detectable improvement when market-state and bond-set controls are shared. Repricing forecasts have lower mean absolute errors than the market-state benchmark in both channels, while incremental ranking evidence is more limited. Trading-attention and credit-direction evidence is weaker. Practical implications The framework can support event screening, analyst prioritisation and issuer-level bond monitoring. Originality/value LLM-assisted preparation enables heterogeneous report information to enter financial research at scale. Scores estimated from historical market outcomes are evaluated out of sample against specified information benchmarks.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: