Why does a backtest lie when later facts are allowed into an earlier day? This note explains point-in-time data, using membership in the S&P 500 as the public example, and states what a clean replay still does not prove.
The day the fact became knowable
A backtest walks through past days and asks what a rule would have done. The walk is only as honest as the information available on each day. Point-in-time data means that a row for a given date contains only what a careful reader could have known at that date, including the delays, restatements, and name changes that arrived later. If today's corrected file is poured backward over last year, the walk sees the correction before the market did. The result can look calm and informed. It is informed by the future.
The failure is easy to miss because the file looks complete. Vendors and public archives often store the latest description of a company, the latest sector, and the latest identifier. A download performed this morning is a snapshot of this morning. Using that snapshot as if it had been sitting on the desk last March is the leak. The leak is not a matter of opinion about models. It is a mismatch between the clock on the row and the clock on the fact.
Point-in-time data is a record in which every fact is stored with the date it became knowable, so a replay cannot see a correction, a membership change, or a new name before that date.
An index membership that moved
The S&P 500 is a public list of companies, and the list changes. A company that is a member today was not always a member. A backtest that builds a portfolio of current members, then plots that portfolio through prior years, has studied a club that did not exist in that form. The companies that later entered were, on the earlier dates, outside the list. The companies that later left were, on the earlier dates, inside it. Treating the current list as a historical list selects survivors and later additions with information no one had.
A point-in-time version of the same study stores the membership as of each date. On the morning a company is added, the row may join. On the morning a company is removed, the row leaves. Announcements often precede the effective date, so the record should also say whether the study is allowed to react on the announcement or only on the effective change. Those are different clocks. Mixing them lets the backtest trade the addition before the rule being tested would have been allowed to know it.
Restated figures create the same leak in a quieter way. A filing can revise a prior quarter. A database that keeps only the revised number will show the revision on the old date. A database that keeps the original print and the later revision, each with its own arrival time, lets the walk see the original first. The S&P 500 example is about membership rather than accounting, but the discipline is the same. The fact and the date it became available are a pair. Storing one without the other is how the lie enters.
What an honest replay still is not
Even a careful point-in-time file can be wrong in ordinary ways. A source can publish late. A corporate action can be mapped to the wrong identifier. A researcher can join two tables on a key that is unique today and was not unique then. The check is to pick a known change, such as a well-documented addition to the S&P 500, and ask whether the study's membership flips on the intended day and not before. If it flips early, the join is using a later file.
Passing that check means the backtest is no longer peeking at that particular fact. It does not mean the rule is sound, that costs are realistic, or that the same rule would have been chosen without looking at the path. Point-in-time data removes one family of lies. It does not bless the idea being replayed. The written record of a study should name the as-of clock in the same place it names the rule.
This piece does not claim that any backtest on the desk is free of look-ahead, that index changes are a strategy, or that a point-in-time file is a complete description of the past. It explains the clock a replay has to respect.
