AI look-ahead bias: why AI can't be trusted to backtest the past

Look-ahead bias is using information in a backtest that wasn't available at the time. The classic versions are mechanical: trading on today's close at today's open, or using earnings figures before they were published. Large language models add a quieter version. A model is trained on text up to a date, its training-data cutoff. Everything written before that date, including every article about how 2020, 2022 or early 2025 turned out, is part of what it learned. So when a model scores a 2022 headline for sentiment (how good or bad the news reads), explains “what drove the 2023 rally”, or picks stocks “as of” some past date, it can't switch that knowledge off.

We use an extreme version of this in our own test suite. A synthetic signal “knows” the next day's return for every date up to the cutoff and flips a coin after it. Before the cutoff it shows a Sharpe ratio above 10 with no drawdown; after the cutoff it looks like what it is, noise. Real leakage is far subtler than that, which is exactly why it is easy to miss: a Sharpe of 1.4 that would have been 0.9 without it doesn't look suspicious.

• A model registry with published cutoffs. Every model the platform calls is listed. Where its training-data cutoff has been read from the provider's documentation, the cutoff is shown with a link to the source. Where it has not, the registry records the cutoff as unknown and the guard assumes the model may know everything up to its release date or, failing that, up to the moment it wrote. A test fails if code starts using a model that isn't in the registry. We use the training-data cutoff, not the earlier “reliable knowledge” date, because the model may know about an event even if that knowledge is patchy.
• Provenance on everything an AI writes. Research answers, deep-research syntheses and backtest verdicts record which model wrote them, its cutoff and when.
• Hindsight labels in research. When an AI-written answer judges a period before the model's cutoff (“did X do well in 2022”, “what caused the SVB collapse”), it carries a label saying it is an explanation written with hindsight, and no evidence that a strategy would have worked. Data answers (prices, fundamentals, macro series from our database) never get the label, even when an AI writes the sentence around them. The check is deterministic: dates, years, quarters, relative phrases like “last year” and a short list of named events such as “during COVID”.
• A backtest guard for AI-derived signals. A strategy input marked as coming from an AI model is split at the model's cutoff (the last day of the cutoff month). Only the window after it can be the headline. The earlier window is shown separately and marked contaminated. If the clean window is shorter than the result card's usual minimum (60 daily or 6 monthly bars), no headline is shown at all. A signal type that looks AI-derived but doesn't name its model is refused, and so is a model we have no verified cutoff for.
• No AI-derived strategies yet. Today's strategies use price and fundamental data only. Ask the strategy agent for a news, sentiment or AI-picked strategy and it says it isn't available yet, and why this matters.

Where a cutoff reads “not recorded”, the registry holds no cutoff for that model yet. The guard then draws the line at the latest date the model could know about: its release date or, failing that, the moment it wrote.

• The cutoff is a fuzzy line. Providers publish it to the month, and a model's knowledge thins out towards it rather than stopping cleanly. We draw the line at the end of the published month; that is a convention, not a measurement.
• General knowledge still leaks. Even for dates after the cutoff, a model carries beliefs formed by what it read before: which companies survived, which themes paid off, how the last rate cycle went. Those priors shape its judgement in ways no date split can remove.
• The clean window is short. It only covers the time since the model's cutoff: a little over a year for the models with a recorded cutoff above, at the time of writing (September 2026). Any AI-derived backtest has wide error bars. The guard makes the number honest; it can't make it precise.
• The research label is a rule, not a reader. It spots dates and question types from wording. An unusual phrasing can slip past it, fiscal years are treated as calendar years, and a question with no date gets no label even if the answer leans on the past.
• Cutoffs come from the provider. If a provider's published date is wrong, our boundary is wrong. Where no cutoff is published we assume the model may know everything up to its release date.