Vendor Restatements and Data Revisions

Data vendors regularly go back and revise historical records — a company restates last quarter's earnings after an accounting error is found, a GDP estimate gets revised as more source data comes in, or a corporate action is corrected weeks after the fact. If you pull "historical" data today, you're typically getting the restated version, not the numbers that were actually known and tradeable on that historical date.

This matters enormously for backtesting: a strategy trading on an earnings-based signal that uses today's restated earnings figure is quietly using information that wasn't available to a real trader on that date, inflating backtest performance in a way that can't be replicated live. The fix is point-in-time (PIT) data — a vendor feed that preserves every version of every data point along with the timestamp it was known, so a backtest can reconstruct exactly what information existed on any given day.

For example, a company reports quarterly EPS of $1.00, then restates it to $0.85 two months later after an accounting review. A non-PIT database simply shows $0.85 for that quarter forever; a PIT database shows $1.00 as the value known during the two months before the restatement, then $0.85 afterward — letting a backtest use whichever value was actually knowable at each historical date.

Ignoring restatements is one of the more insidious forms of look-ahead bias, because the affected numbers look completely reasonable and the error doesn't announce itself in any error log.