Every historical dataset tells two stories. One is the story of the market. The other is the story of the record-keeping that survived to the present. A model trained on the second is being taught hindsight and graded on it — often without anyone having chosen the deception.

The earlier essays in this library looked at how simulation and leakage flatter a strategy. This one is about the substrate: the historical record itself, and what it takes to read it as a person standing in the past actually could.

The default dataset is written backwards

Load a table of index constituents today and you receive the membership of the present, stamped retroactively into history. The firm that was removed after its collapse is missing precisely where you needed to see it. Delisted equities, dissolved funds, discontinued series — all are curated out by the simple act of looking backwards from now.

This is the graveyard problem, and it is a data pipeline concern before it is a statistics concern.

Restatements: history edited in place

A delicate and underrated hideout. Regulatory filings are amended; accounting standards change; dividends are reclassified. Databases handle this by overwriting the record — the revision replaces the release. A model studying 2019 with post-restatement numbers experiences a cleaner world that 2019 never was.

Point-in-time data archives the release, not the revision: every figure is stored with the date it became public, and analysis proceeds as if future corrections had not been written yet. The model then faces the same fog a real analyst faced.

Reconstitution, and the sum of vanishing edges

Indices are revised quarterly: names enter on the way up and leave on the way down, and benchmark history silently chants the winners. Timing and group membership effects are perhaps the most widely used “features” in quantitative finance, and the difference between as-of and as-today membership flows directly into an evaluation.

Ticker recycling adds a comedy of its own: a symbol inherits the past of a defunct company, merging two companies into one ghost.

The as-of join as a habit, not a feature

Mechanically, point-in-time discipline is unglamorous — a bigint of timestamps and a rule for joining: every record is paired with the latest earlier version the market had seen, never the closest known today. It is diligence, and it costs real engineering: point-in-time archives are expensive to maintain, which is precisely why the shortcut is so common and so profitable to correct for.

What honesty changes downstream of it

On a point-in-time substrate, an index-tracking strategy earns a lower, truer number: entries known in advance, deletions processed on public schedules, amended figures unseen until their release. The same evaluation on a convenient dataset inflates, and by a widening band over longer horizons.

Backtests on a hindsight substrate are trained on a stage play about the past; point-in-time data puts the real archive back under the model — revisions pending, corpses included, exactly as they were.

Frost Investment Research publishes free editorial essays about how machine learning gets evaluated in investing. Nothing here is personalised advice, and nothing on this site is for sale — including this line of work. Spotted an error, or want the next essay to cover something specific? Tell the editorial desk.

← Back to the article library