Point-in-time, or it didn't happen.

A backtest can only be as honest as the data it sees. The most common way research goes wrong is not a bad model; it is a dataset that quietly knows the future.

Two clocks, not one

Every observation carries two timestamps: the time it describes and the time it became knowable. A quarterly earnings figure describes the quarter that ended in March, but the market learned it in late April — and it may be restated in August. A point-in-time store keeps every version, stamped with when it was published.

If a value was not knowable at time t, a strategy evaluated at time t must not see it.

Where look-ahead bias hides

• Restatements overwriting the original release.
• Index membership taken from today's constituents instead of the list on each date (survivorship).
• Corporate actions applied to prices before the announcement date.
• Bar timestamps marking the open of a bar whose values are only final at its close.

What we do about it

Our ingest layer writes append-only, versioned records with an as_of timestamp. Research queries take an as_of parameter and can never read past it. The same query run tomorrow returns the same answer — which is the whole point.