Lookahead bias: the bug that makes backtests lie — and how our own got caught

Lookahead bias is the backtest's native sin: any pathway by which information from the future leaks into a decision made in the past. The result is a simulation that could not have been traded, printing profits that were never available. Every backtesting practitioner knows the definition. QuantLab still shipped a case of it — caught by its own audit, months in, on its proudest result. This is the definitive explanation, from inside the incident.

The everyday disguises

The obvious version is trivial: deciding to buy at today's open using today's close. The dangerous versions wear suits. Normalization leakage: scaling features with a mean and deviation computed over the whole dataset, including the future. Selection leakage: choosing the strategy because it won on data you then “validate” on. The exit-anchor: computing a decision's features at the time the trade closes, while simulating it as if known at entry. That last one is the one that got us.

The incident, plainly told

The lab's trend strategy carried a machine-learning condition gate — a filter that decides whether a trade is worth taking. On paper the gated strategy was the campaign's champion: logged backtests around a $100 start simulating to roughly $199, profit factors comfortably above 1.3 across years. Then the T34 audit asked one hostile question: at which bar are the gate's features actually computed? Answer: the exit bar. The gate was reading the trade's outcome while “deciding” whether to take it — the future, piped politely into the past. Re-gated honestly at the entry bar, the same champion simulated to roughly $106 — a coin flip after costs. The entire premium was the bug. The numbers are framed here as what they now are: historical backtest results that did not survive re-testing.

What the lab changed

The audit that caught it (documented in the repo's exit-model audit) became mandatory: every promoted result gets a causal audit — prove, feature by feature, that each input is computable strictly before the decision it feeds. The deeper defenses are structural: point-in-time feature construction in the shared engine, gates anchored to the entry bar by construction, and a culture where re-deriving a result with an independent script is expected, not paranoid. The project had already retracted one result batch for a related measurement artifact; this audit exists because of it.

The transferable lesson

Lookahead bias is not a typo you make once; it is the default direction of every mistake when you write a simulator quickly, because the data is all sitting there in one frame and the future is the easiest column to reach. The cure is not caution — it is architecture and audit: one engine that makes leakage hard, and a standing rule that your best result is the one that gets the most hostile re-examination. The broader defense-in-depth is the full validation gauntlet; the taxonomies of failure it guards against are in why backtests fail.

Research note: QuantLab experiments are presented for educational and research purposes. Historical backtests and simulations do not guarantee future results and should not be interpreted as investment advice.

Sources

• Ojeology/QUANTLAB — EXIT_MODEL_AUDIT.md: the exit-engine correction that superseded R066–R072
• The blind-validation branch — T34 entry-anchor audit documentation