加载中...
本论文提出Uni-FinLLM,一种将价格时序、金融文本、宏观指标和图结构信息通过共享Transformer骨干与模块化任务头联合建模的多模态LLM,旨在同时实现微观股票预测与宏观系统性风险预警。实证结果显示:在微观层面将方向准确率从61.7%提升至67.4%,MAPE降至10.9;在信用风险预测中Accuracy达84.1%、ROC-AUC达0.892;宏观预警准确率提升至82.3%(危机F1=79.8%),验证了统一多模态表示与跨尺度任务共享的有效性
Table 1. Micro-Level Stock Prediction Performance
| Model | Directional Accuracy | MAPE | Hit Ratio |
|---|---|---|---|
| ARIMA (baseline) | 53.2% | 18.7 | 50.4% |
| LSTM | 57.8% | 15.4 | 54.2% |
| Temporal Fusion Transformer | 60.1% | 13.8 | 56.9% |
| FinBERT (text only) | 58.6% | 14.7 | 55.1% |
| Llama-Fin | 61.7% | 13.2 | 57.6% |
| Uni-FinLLM (ours) | 67.4% | 10.9 | 64.3% |
Table 2. Credit-Risk Prediction Performance
| Model | Accuracy | F1-Score | ROC-AUC | PR-AUC |
|---|---|---|---|---|
| Logistic Regression | 71.3% | 69.1% | 0.742 | 0.701 |
| Random Forest | 74.5% | 73.4% | 0.781 | 0.731 |
| FinBERT | 77.2% | 76.8% | 0.812 | 0.784 |
| Llama-Fin | 79.6% | 78.4% | 0.836 | 0.813 |
| Uni-FinLLM (ours) | 84.1% | 83.6% | 0.892 | 0.871 |
Table 3. Macro Systemic-Risk Early-Warning Performance
| Model | Accuracy | F1-Score | ROC-AUC |
|---|---|---|---|
| Logistic Regression (SRISK inputs) | 68.5% | 64.1% | 0.743 |
| VAR Model | 70.4% | 66.5% | 0.762 |
| LSTM-Macro | 72.8% | 68.7% | 0.789 |
| GNN-MacroRisk | 75.6% | 72.4% | 0.816 |
| Uni-FinLLM (ours) | 82.3% | 79.8% | 0.873 |
元数据与概览(引言与报告概览)
逐节深度解读
图表深度解读(重点图表与表格逐一解析)
估值/经济价值类分析(模型未涉)
风险因素评估(报告识别与外部评估)
批判性视角与细微差别
结论性综合
(若需,我可以继续:1) 提取并整理报告中所有数学符号与公式的逐项解释;2) 根据表中指标为每个基线计算性能差距的置信区间建议检验方案;3) 列出对监管采用该系统的合规性与可审计性改进清单。)
2601.02677.pdf