Why Point-in-Time Data Matters in Research and Backtests

Point-in-time data means using the version of a number that was actually knowable on the date you claim to have known it, not the version sitting in your database today. It matters because companies routinely restate their own prior-year figures, so testing an idea on “as reported today” history can quietly make your research and your backtests look smarter than they ever were.

Here is the trap in one line. The number you pull for a past quarter is often not the number that was on the tape that quarter. And if your analysis does not respect that difference, it is borrowing information from the future without meaning to.

The number changes after the fact

When a company reports a quarter, it does not just print this period’s results. It also prints the comparable period from a year ago, so you can see the growth. Most of the time that prior figure matches what was originally filed. But often it does not, and for good reasons.

A company restates its own history when the basis of the numbers changes. The usual causes are well known and entirely routine:

• Accounting standard changes. When a new rule on leases, revenue recognition, or financial instruments comes in, prior years get recast onto the new basis so the trend is comparable.
• Demergers and spin-offs. When a business is hived off into a separate listed company, the parent’s earlier revenue and profit are restated to exclude it, so you are comparing like with like.
• Discontinued operations. When a company decides to sell or wind down a division, that division’s results are pulled out of continuing operations and shown separately, including in the prior-year column.
• Segment redefinitions. When management reorganizes how it reports its businesses, the historical segment split gets redrawn to match the new structure.

None of this is a scandal. It is the correct thing to do. A serious analyst wants comparable periods, and restatement is how you get them. The problem is not that companies restate. The problem is what happens when your data provider silently overwrites the old figure with the new one, and your research then treats the new figure as if it had always been visible.

Why “as reported today” flatters research

Most financial databases are built to answer one question: what are this company’s numbers for this period. They are not built to answer the question that actually matters for testing an idea: what did we know, and when did we know it.

So when you download a company’s ten-year history today, you often get the latest, cleanest, most restated version of every line, stamped against the old dates. The demerged business has already been stripped out of the years before the demerger. The discontinued division has already been separated from continuing operations. The new segment map has already been applied backward.

That tidy history is exactly what fools you.

Imagine an analyst building a case that a company’s core business was quietly accelerating three years ago, before the market noticed. If the “core” they are measuring is a clean, continuing-operations figure that only exists because a division was later carved out, they are describing a business that did not exist in that shape at the time. The signal they think they found was manufactured by a restatement that happened afterward. The analysis is not dishonest. It is just quietly looking backward through a lens that was ground later.

The same trap catches models. This is the mechanism behind lookahead bias, where a test uses information it could not have had on the day it claims to act. Restated comparables are one of the cleanest ways lookahead sneaks in, because nothing looks wrong. There is no obvious error, no missing file, just a number that is a little too good because it was tidied up after the fact.

A concrete way to feel it

Think about a company that sells off a low-margin division. Before the sale, its reported margins include that drag. After the sale, and in every restated prior year, the division is gone and the historical margins look structurally higher.

Now run a naive study: “companies whose margins were above X three years ago went on to outperform.” If your data is restated, some of those high past margins are an artifact of divisions that were only removed later. Your study has quietly selected for companies that would go on to clean up their portfolios, which is a fact from the future dressed up as a fact from the past. The rule looks predictive. It is partly just remembering the answer.

This is why the same discipline that protects a backtest also protects ordinary research. It is not only quant desks that get burned. Any analyst comparing “then” to “now” using today’s recast history is exposed.

What disciplined desks actually do

The fix is conceptually simple and operationally unglamorous. You store data the way it arrived, not the way it ended up.

That means keeping the figure as it was originally filed, with the date it was first public, and keeping each later restatement as a separate record rather than painting over the old one. When you then ask, “what did this company’s history look like as of a date three years ago,” you get the version that was visible then, not the version that exists now. When you want the clean restated series for a fair long-run trend, you can have that too. The point is that you can tell the two apart and choose deliberately.

A few habits fall out of this:

• Separate the period from the knowledge date. Every value should carry both when it describes and when it became knowable. A March quarter is not public in March.
• Never overwrite a restatement. Keep the original and the revision side by side. The gap between them is often the most interesting thing in the file.
• Ask which basis you are on. Before comparing a company to its own past, check whether a demerger, a discontinued operation, or a segment redraw sits between the two dates. If it does, the raw comparison is not apples to apples.

None of this makes for exciting work. It makes bigger databases and more careful queries. But it is the difference between research you can defend and research that is quietly lying to you.

The judgment layer, not just the plumbing

There is a second, subtler point that separates good analysts from good data. Spotting a restatement is not the same as understanding it. A restatement is a signal that the shape of the business changed, and the right response is to go read why.

This is where reading the annual report earns its keep. Demergers, discontinued operations, and segment changes are all explained in the notes and the management discussion. The number moving is a prompt. The explanation is where the actual insight lives: was a division sold because it was a drag, or because it was worth more to someone else? Did segments get redrawn to clarify the business or to bury a weak one inside a strong one?

This is also the honest boundary for AI in research. A model is very good at flagging that a prior-year figure changed between two reports. It is far weaker at knowing whether that change should alter your view, and it is dangerous if it silently trains or tests on the restated version. As we have argued in data quality beats model quality, the cleverest model in the world is only as honest as the history you feed it. If that history is the polished, future-informed version, even an AI that could otherwise predict earnings is grading itself on an exam it has already seen.

The most dangerous data is not the data that is wrong. It is the data that is right today and was not knowable then.

The quiet discipline

Point-in-time data is one of those foundations that gets no applause when it is done and causes no obvious alarm when it is skipped. That is exactly why it is so easy to neglect. The results of neglecting it do not look like errors. They look like slightly better numbers, a slightly sharper edge, a strategy that seems to work until real money is on it.

Companies will keep restating their histories, and they should. The work on your side is to remember that the past you can see today is not always the past that was visible then, and to build every study, every screen, and every comparison on the version of history that was actually knowable at the time. Get that right and everything above it has a chance of being true. Get it wrong and you are, very politely, marking your own homework.