Hands-On Machine Learning for Algorithmic Trading

Look-ahead bias

Tests of trading rules derived from past data will yield biased results when the sample data used to develop the rules contains information that was not, in fact, available or known at the point in time the data refers to.

A typical source of this bias is the failure to account for the common ex-post corrections of reported financials. Stock splits or reverse splits can also generate look-ahead bias. When computing the earnings yield, earnings-per-share data comes from company financials with low frequency, while market prices are available at least daily. Hence, both EPS and price data need to be adjusted for splits at the same time.

The solution lies in the careful analysis of the timestamps associated with all data that ...