17 published LLM trading papers: how much of each backtest did the model not already know?17 篇公开的 LLM 交易论文：每个回测里，有多少个月是模型没读过的？

Cutoffs are taken only from the model provider's own documentation, never from the paper. Where none is published, the row says so. Facts only — no judgement of the papers' quality or intent.截止日只取模型方的官方文档，不取论文自报。查不到的如实标注。只陈述事实，不评价论文质量或作者意图。

48 months is what a Sharpe-1 claim needs to clear t = 2 (Lo 2002, i.i.d. — a floor). At Sharpe 2 the bar is 12 months, and two papers would clear it.48 个月是夏普 1 的主张达到 t = 2 所需的干净月数（Lo 2002，假设 i.i.d.，是下限）。若按夏普 2，门槛为 12 个月，有 2 篇够长。

1 · Nine papers with a documented cutoff一 · 查得到官方截止日的 9 篇

What the papers say, quoted论文原话

• CryptoTrade — "We carefully select the test period after September 2021, the GPT-3.5's knowledge cutoff date". The headline runs use GPT-4 and GPT-4o. Test segments are chosen by market label (bear, sideways, bull).「We carefully select the test period after September 2021, the GPT-3.5's knowledge cutoff date」；主实验用的是 GPT-4 与 GPT-4o。三段测试期按熊、横、牛行情标签选取。
• Expert Investment Teams — "GPT-4o … with the knowledge cutoff date in August 2023". The provider page says October 2023, which puts September and October 2023 inside the test.「GPT-4o … with the knowledge cutoff date in August 2023」；官方页面写 2023 年 10 月，按官方则 2023 年 9、10 月在测试期内且已读。
• Strat-LLM — "Operating in a live-forward setting throughout 2025 … to eliminate look-ahead bias"; "T+1 rolling-window backtest using real-time 2025 data". The headline model's documented cutoff is 2025-08-31.「Operating in a live-forward setting throughout 2025 … to eliminate look-ahead bias」「T+1 rolling-window backtest using real-time 2025 data」；主模型官方截止日为 2025-08-31。
• EvolveTrade — chooses windows "after each model's knowledge cutoff" — the right design; the windows are one to three months long.按「after each model's knowledge cutoff」选窗口，设计正确；窗口长 1 到 3 个月。
• FinMem — the provider page lists the current GPT-4-Turbo snapshot; an earlier snapshot can only have an earlier or equal cutoff, which leaves the test entirely before it either way.官方页面列的是现行 GPT-4-Turbo 版本；若论文用的是更早版本，截止日只会更早或相同，测试期仍全部在截止日之前。

2 · Eight papers that cannot be assessed二 · 无从判断的 8 篇

3 · What the two tables say三 · 两张表说明了什么

1. None of the nine reaches 48 clean months.9 篇里没有一篇的干净样本达到 48 个月。
2. Four tested mostly on months the model may already have read.4 篇的测试期大半落在模型可能读过的月份里。
3. Five of the eight that cannot be assessed used open-weight models that publish no cutoff at all.无从判断的 8 篇里，5 篇用的开源权重模型根本没有公布截止日。
4. Since 2025 most papers test after the cutoff. The problem is moving from open-book to too short.2025 年以后多数论文已在截止日之后测试，问题正从「开卷」变成「太短」。

Prior work前人工作

This is not a new criterion. The Alpha Illusion lists model version and knowledge cutoff among minimum disclosures; What survives honest evaluation? works the same power arithmetic; Glasserman & Lin (2023) measured look-ahead bias in GPT sentiment signals. What this page adds is narrower: a paper-by-paper table with each cutoff traced to the provider.这不是新判据。Alpha Illusion 把模型版本与知识截止日列为最低披露；What survives honest evaluation? 做的是同一种样本量算术；Glasserman & Lin（2023）最早量化了 GPT 情绪信号的前视偏差。本页的增量更窄：一张逐篇、每个截止日都追溯到模型方的表。

Method and limits做法与限定

• 105 candidates from five arXiv API queries; 18 selected as LLM trading frameworks reporting historical backtests; 1 excluded (planned experiments, no results). Full texts read 2026-10-05.五组 arXiv 检索得 105 篇候选，筛出 18 篇报告历史回测的 LLM 交易框架，排除 1 篇（只有实验计划）。全文于 2026-10-05 读取。
• Computed with effective_window from this repository: the cutoff month counts as read; required length for Sharpe 1 at t = 2.用本仓库的 effective_window 计算：截止当月算已读；所需长度按夏普 1、t = 2。
• "Before the cutoff" means the model may have read it, not that it did.「截止日之前」指无法排除读过，不是一定读过。
• Headline results only; one auditor; selection may have missed papers.只看主结果；一人审计，候选筛选可能有遗漏。
• Corrections welcome — and so are additions: three dates and a trial count are enough.欢迎指正，也欢迎补充：三个日期加一个试验次数就够。

Evidence Deflator · GitHub · pip install evidence-deflator