Abstract
This study developed an end-to-end framework for identifying asymmetric financial risk factors by combining LoRA enhanced Large Language Model (LLM) embedding with risk sensitive causal forest. This method first converts enterprise level financial news into domain adapted semantic covariates, and then estimates heterogeneous downward processing effects under explicit causal identification assumptions. The splitting criterion adds asymmetric penalties on the basis of the standard causality forest heterogeneity objective, giving greater weight to negative treatment effects and fat tail bias. This framework utilizes approximately 1.2 million daily stock observation data and is evaluated using out of sample testing, MSE conversion results, Qini/AUUC ranking metrics, permutation based robustness checks, and component level ablation specifications. The MSE of this model is about 0.40, which is 15.8% higher than the DML and GRF benchmarks. The top 10% of risk quantiles have increased by >3.5 times, and the top 20% of samples cover over 60% of cumulative downside risk. The explanation based on SHAP further indicates that negative news sentiment is the main semantic driving factor of asymmetric risk, especially in the banking industry, where the impact of bad news shows a significant downward amplification effect.