加载中...
本文提出 Look-Ahead-Bench:一个用于评估金融场景中 Point-in-Time(PiT)大模型“前视偏差”(look-ahead bias)的标准化基准,采用双期(in-sample / out-of-sample)真实交易工作流与 alpha decay 指标衡量模型记忆性与泛化能力,实验证明常规模型在迁移到后验期时存在显著 alpha 衰减,而 PiT 模型在去除未来记忆后随规模提升表现出更好的推理能力与更小的 alpha decay。
| Model / Strategy | Variant | P1 Return (%) | P1 Alpha (pp) | P2 Return (%) | P2 Alpha (pp) | Alpha Decay |
|---|---|---|---|---|---|---|
| QUANTSTRATEGIES(Baseline) | ||||||
| Buy & Hold | Passive | +25.32 | +0.00 | +24.75 | +0.00 | +0.00 |
| Equal Weight | Systematic | +25.68 | +0.36 | +22.33 | -2.42 | -2.78 |
| Momentum (3M) | Systematic | +33.28 | +7.96 | +30.50 | +5.75 | -2.21 |
| Mean Reversion | Systematic | +20.70 | -4.62 | +9.35 | -15.40 | -10.78 |
| MA Crossover | Trend | -2.46 | -27.78 | +6.91 | -17.84 | +9.94 |
| Random Noise | Control | +11.43 | -13.89 | +0.56 | -24.19 | -10.30 |
| AI MODELS (Agents) | ||||||
| Llama 3.1 8B | Standard | +39.13 | +13.81 | +21.33 | -3.42 | -17.23 |
| Llama 3.1 70B | Standard | +44.59 | +19.27 | +28.77 | +4.02 | -15.25 |
| DeepSeek 3.2 | Standard | +46.05 | +20.73 | +23.71 | -1.04 | -21.77 |
| Pitinf-Small | PiT | +25.07 | -0.25 | +24.81 | +0.06 | +0.31 |
| Pitinf-Medium | PiT | +27.76 | +2.44 | +28.04 | +3.29 | +0.85 |
| Pitinf-Large | PiT | +31.34 | +6.02 | +32.07 | +7.32 | +1.30 |
元数据与概览(引言与报告概览)
逐节深度解读
方法论详解(Methodology)
图表深度解读(重点:Table 1)
对 Table 1 的逐行解读与洞见:
AI 模型(代理)部分:
对 Table 1 的方法学注释与潜在局限:
估值/模型评估方法(报告中未包含传统估值)
风险因素评估(报告识别与建议)
批判性视角与细微差别
结论性综合
总体评价(要点摘录)
2601.13770.pdf