Restatements and Data Revisions
Reported financial and economic figures get corrected or restated after the fact, sometimes materially — a backtest that uses only the final, cleaned-up numbers is testing a strategy against information that traders at the time never had.
Prerequisites: Point-in-Time Data
A company's reported earnings, GDP growth, employment figures, and even prior-quarter revenue are not fixed once published — they get revised, sometimes repeatedly, sometimes substantially, often months or years later as more complete information comes in or as accounting corrections are made. A backtest that pulls "historical" data from a vendor is usually pulling the most recently revised version of every past number, not the version that was actually publicly known on that historical date. That gap is a direct source of look-ahead bias.
The idea
Revisions come from a few common sources: initial economic releases (like nonfarm payrolls or GDP) are estimates built from incomplete survey data and get revised as more responses come in over following months; corporate earnings get restated when accountants find and correct errors, sometimes years after the original filing; and index-level data gets recalculated when its constituent methodology changes retroactively. In each case, the number a database shows you for "March 2019" today can differ meaningfully from the number that was actually reported and traded on in March 2019.
The practical fix is to always use point-in-time or vintage data (data stamped with the date it was actually published) for anything that feeds a trading signal, rather than a "history" table that silently reflects only the latest revision.
Worked example
US GDP growth for a given quarter is typically published as an "advance" estimate about a month after quarter-end, then revised twice more over the following two months as more complete trade and inventory data arrives, and can be revised again years later during annual benchmark updates. A strategy that trades on GDP surprises — comparing the released number to consensus forecasts — must compare the forecast to the advance estimate that traders actually reacted to on the release date, not to the fully revised figure a database shows for that quarter today; using the revised figure would substitute a number nobody could have known at the time for the one that actually moved markets.
What this means in practice
Any backtest using macroeconomic releases, corporate fundamentals, or analyst estimates needs to confirm explicitly whether the data vendor is providing point-in-time vintages or only the latest values, since the two can produce very different — and misleadingly good — backtested results for signals based on "surprises" relative to what was known at the time.
Reported financial and economic data gets revised after publication, sometimes substantially — a backtest must use the value as it stood on the historical date in question, not the fully revised figure a database shows today, or it will be silently trading on information from the future.
Further reading
- Diewert, 'The Economic Theory of Index Numbers and the Measurement of Prices and Real Output', ch. 2