Look-Ahead Bias

What look-ahead bias is

At any real moment, a bot knows the past and nothing else. A backtest can break that by accident. It holds the whole history in memory, so a rule can reach for a value that, on that day, had not happened yet. The test then trades on the future.

The results look great because they are great, in the one place that does not exist: the test. Live, the future is not available, and the edge goes with it.

The four common types

• Using the close before the bar closes. A rule says "enter if the bar closes above the average" and the test enters at that bar's open or at a price inside the bar. At the open, nobody knows where the bar will close. The test is trading with the answer.
• Using the high or low of the current bar. "Buy at the lowest price of the bar" or a stop placed using the bar's high needs the whole bar. Live, the low is only known after the bar ends.
• Indicators that repaint. Some indicators change their past values as new bars arrive, such as zigzag and swing-point tools that only confirm a turning point later and then redraw it as if it were known earlier. In a backtest the finished picture is used for every bar, so every turn looks perfectly caught.
• Future data in normalisation. Scaling a series with its overall minimum, maximum, mean or standard deviation uses numbers that depend on later data. The scaled value at bar 100 depends on what happened at bar 1,000.
• Adjusted and restated data. Prices adjusted for splits and dividends, or economic figures later revised, show values that were not what traders saw at the time. A rule built on the final revised number trades on information that arrived weeks later.

A worked example: same bar close

Here is a rule with no edge at all: "if the bar closes higher than it opened, go long." Eight made-up bars, with the open and close of each:

All prices are made up. The peeking rule decides using the close and then trades from the open of the same bar. It can only ever trade bars that went up, so it takes 5 trades, wins all of them and makes +7.0 points. The honest rule can only act after the bar has closed, so it enters at the next bar's open and has to take whatever that bar does. It takes 4 trades and makes -3.5 points. In this tiny sample, the "edge" was entirely the peek.

The same thing happens with highs and lows in less obvious ways. If a rule says "buy at the low of the bar when price touches the average", the test fills at the best price of the bar by definition. No live order is assured that fill.

How it hits bots

• Grid bots: a backtest that assumes each grid level fills at exactly the level price whenever the bar's range touched it uses the high and low of the bar. It cannot tell whether the low or the high came first, and that order decides whether the bot was in or out.
• Martingale bots: if the "take profit hit" check uses the bar's high and the "stop" check uses the low, the order of events inside a bar matters. A test that assumes profit first ignores bars where the loss came first.
• DCA bots: buying at the day's low or a weekly "dip" identified with the full week's data peeks at the future.
• EAs: a published EA that opens on "bar 0" and uses indicator values from bar 0 can trade on a candle that is still moving. Results on a completed bar differ from results while it forms.

How to check for it

• Run bar by bar. Feed the rule one bar at a time and let it see only what existed. A rule that cannot run this way is a warning.
• Delay every signal by one bar. If the result collapses, as it did above, the rule relied on information from the same bar. Some real rules lose a little from a delay, and that is fine. A collapse is the sign.
• Compare backtest and live signals. Record the live signal and rerun the same time period in the backtest. The two should match, bar for bar. A signal that appears in the backtest and never appeared live was drawn after the fact.
• Look at the indicator's rules. Does it change old values? Does it confirm turning points with later bars?
• Check the data source. Is the data point-in-time, as it was known then, or adjusted or revised later?
• Be suspicious of perfection. A smooth curve with a very high win rate and no deep drawdown is often a data problem before it is a good strategy. See overfitting for the other common cause.

What it does not tell you

Passing the delay test does not mean the strategy has an edge. It means the test was not peeking in this particular way. Other biases remain, such as survivorship bias in the data and fitted settings. Look-ahead bias is also not always an obvious bug: a few lines of code can hide it, and a trading platform's built-in tester can hide it in how it models the current bar. Read how your tester handles the open bar, and run the delay check on anything you rely on.

Finally, the only complete test is forward testing on data that did not exist when the rule was written.