A backtest is a claim about the past: given only what was knowable on this date, the rules would have done this. Look-ahead bias breaks that claim by letting the simulation read something from the future.
It is rarely deliberate and almost never obvious. The usual result is a beautiful equity curve that cannot be reproduced with real money.
The four common leaks
Financial statements dated to the period rather than the filing. Under Regulation 33 of SEBI’s LODR Regulations, a listed company files quarterly results within 45 days of quarter end, and annual results within 60 days (SEBI FAQs on LODR). Ranking stocks on 30 September using the September-quarter numbers means trading on figures nobody had until November.
Index membership taken from today. Backtesting a Nifty 50 rotation on the current 50 names assumes you knew in 2015 which companies would still be in the index in 2026.
Same-bar execution. Deciding on a bar’s close and filling at that same close is physically impossible, since you are still observing the bar when you act.
Restated or back-adjusted series. A price history adjusted for splits and bonuses today is not the price you saw at the time, and a restated financial is not the number that was originally filed.
What honest handling looks like
| Leak | Correct treatment |
|---|---|
| Fundamentals | Use the publication date, not the period end date |
| Index or universe membership | Rebuild the constituent list as of each rebalance date |
| Execution timing | Decide on one bar, fill on the next bar’s open |
| Corporate actions | Apply the adjustment only from the ex-date forward |
Orders in a saral.money backtest fill at the next bar’s open, which closes the execution-timing leak by construction.
Caveats
Look-ahead bias is hard to detect from the outside, because a strategy with a leak looks like a strategy that works. The tell is a result that is too clean: a curve with shallow drawdowns, an unusually high hit rate, and performance that holds up in periods when the market itself did not.
When a backtest looks unusually good, audit the data path before you celebrate the idea. The most likely explanation is that a date is wrong somewhere.