The Qadira value backtest uses book value from annual reports dated to the fiscal year-end, although reports appear about 60 days later. The backtest earns 11% a year. What bias is this, how does it inflate returns, and how do you fix it?

This is look-ahead bias: the backtest trades on book values about 60 days before they were published. Stocks whose annual numbers turn out well tend to rise in those 60 days, and the backtest buys them before the move. Lagging book value to its release date cuts the return from 11.0% to 7.9%, so about 3.1 points, 28% of the return, was never available. Fix it with point-in-time data stamped by filing date.

Think of betting on a cricket match after reading the score in tomorrow's newspaper. The score is accurate; the problem is when you knew it. A backtest may use a number only from the day it was public, and Qadira's book values were stamped with the date they describe, 31 March, not the date they became known, about the end of May. That is {term('look-ahead bias', 'Using, in a backtest, any information that was not available at the moment the simulated trade was made.')}, and it hides inside databases that store figures by the period they belong to.

Between year-end and filing, the market learns a lot about the year: quarterly results, management comments, sector data. Stocks whose annual book value will come in strong tend to rise in that window. The biased backtest ranks on the final numbers and buys on day 1, so it holds those stocks through the rise that the real investor could only have seen afterwards. The short side works the same way in reverse. Every rebalance gets a 60-day head start, and it compounds across 15 years.

Read the lag test. Return falls from 11.0% with no lag to 7.9% at 60 days, then barely moves to 7.6% at 90 and 7.5% at 120; the fall stops where the information became public. So about 3.1 points, 28% of the headline, came from the future. The flat part after 60 days is the reassuring sign: the strategy's remaining return does not depend on the exact lag, which is what a genuine effect looks like.

Store every data point with the date it became available, a point-in-time database, and let the backtest see only what was public at each simulated moment. Use each company's actual filing date where you have it; where you do not, apply a conservative uniform lag, such as 90 or 120 days, and check that results survive it. Two relatives of the same bias need the same care: restated figures, where the database has overwritten the first reported number with a later correction, and index membership, where today's constituents are used for past years. Say the limitation: a lag rule cannot recover information about late filers that the database never recorded.