Lookahead bias: the four ways a backtest sees the future
Lookahead bias is not a bug you notice. The backtest runs, the numbers are plausible, and the equity curve is smooth — because the strategy was quietly allowed to read data that had not happened yet.
1. Negative shifts
signal = df["close"].shift(-1) > df["close"]
The minus sign moves tomorrow's close into today's row. In a live system that value does not exist at decision time. The correct form lags the signal instead:
position = (fast > slow).shift(1).fillna(0.0)
2. Future indexing
if df["close"].iloc[i + 1] > df["close"].iloc[i]: ...
Same defect, different syntax. iloc[i+1], iat(i+1) and
loc[index+1] all reach one row past the decision point. So does
pct_change(-1).
3. Whole-series statistics
z = (px - px.mean()) / px.std()
This is the subtle one. There is no negative sign and nothing looks wrong — but
px.mean() is computed over the entire history, including the future. At row 10 the
strategy "knows" the average of row 1,000. Any normalisation, threshold, or signal built from a
full-sample mean, standard deviation, min, max, median, quantile, skew or kurtosis has the same
problem.
The causal version uses only what existed at the time:
z = (px - px.rolling(60).mean()) / px.rolling(60).std()
Computing avg = returns.mean() once for a Sharpe ratio in your report is fine —
that is a statistic about the result, not an input to a decision. The distinction is
dataflow, not syntax, and it is exactly the distinction a naive scanner gets wrong.
4. Point-in-time data
Fundamentals keyed to the fiscal period rather than the publication date, index membership taken from today's constituents, or a dataset that was revised after the fact. The code can be perfect and the backtest still dishonest, because the data knew the future.
How to catch it
- Read the code for the three syntax patterns above — they are grep-able.
- Perturb the data after time t and re-run: if any output at or before t changes, future information is leaking.
- Shift the entire dataset forward by one bar and see whether performance collapses. A real edge degrades; a lookahead edge disappears.
- Audit it automatically. Adwa's static layer parses the AST rather than matching strings, so it catches the patterns above without flagging honest metric code.