Radar Perene / Archive / science

Look-ahead bias: the error that lets a model \"predict the past\"

◦ Index methodology v2.2 (working papers with DOI). See the methodology.

Science

A student hands in a perfect exam. The teacher grows suspicious — and discovers the answer key was sitting on the desk the whole time. No one would call that talent. In market research, however, the exact equivalent of that episode circulates every day under the name of a result: a model tested on the past using information that, in that past, did not yet exist.

Look-ahead bias is the methodological error of allowing a test on historical data to use information that only became available after the date being simulated. The model appears to predict the future; in fact, it consulted the future to describe the past.

The concept is universal: no conclusion about "what would have worked" is worth anything if the test saw the answer key.

How the answer key slips into the exam

The scandalous case is rare. Almost no one deliberately programs a model that reads tomorrow's price to decide today. The bias enters through quiet cracks, and three of them account for most cases.

The first is the release date. A quarter's GDP is published months after the quarter ends; a corporate balance sheet, weeks after the closing date. A test that uses the data on the date of the fact — rather than the date the data became public — hands the model an advantage no participant in that session ever had.

The second is series revision. A large share of macroeconomic series gets revised: the employment figure released in March is not the one sitting in the official series downloaded ten years later. Whoever tests on the final series is using a version of the past that only existed in the future.

The third is the subtlest: the contaminated decision. The researcher chooses which assets, which windows and which rules to test already knowing how the story ended. No line of code looks ahead — but the entire study design was written by someone who did. This is the crack that connects the error to its cousins, survivorship bias and overfitting: in all three, the past being tested is not the past anyone actually lived.

The antidote has a name: point in time

The defense against look-ahead bias is not intelligence — it is archival discipline. It is called point-in-time data: for every simulated date, the test only sees what was public and available on that date, in that date's version.

It is an expensive discipline. It requires storing not just the values but the calendar of when each value appeared — and resisting the temptation to "fix" the archive when a better version of the data shows up later. An honest archive ages with its own period errors inside it.

This house operates under that rule by construction. Every entry in the Daily and every essay in the Atlas is written only with what that date's close contained — the later outcome does not enter the reading, even when the editor knows it. The episode of the Selic at 3.75% in March 2020 is an example of the genre: the text records what the date knew, not what 2021 revealed. Once published, the record is locked — reprocessing the system does not rewrite the published past. In numbers: the public archive holds around 280 essays built under that rule, and the data package that accompanies the house's research — series, dated events and prior registration of hypotheses — is deposited with a public DOI (10.5281/zenodo.21399426), verifiable by any reader.

Why the error survives

If the antidote is well known, why does the bias persist? Because it is pleasant. A contaminated test produces beautiful curves, and beautiful curves travel better than methodological caveats. Whoever sells a model rarely has an incentive to audit the timeline of their own data; whoever buys one rarely knows that this audit is the first question to ask.

There is also a less cynical reason: verifying the historical availability of every input is tedious work, invisible in the final result. The difference between a study with and without that audit does not show in the chart — it shows years later, when live performance parts ways with the simulation. Look-ahead bias is a debt the test takes out against the future.

Frequently asked questions

Is look-ahead bias the same as overfitting?

No. In overfitting, the model fits too closely the noise of the data it saw; in look-ahead bias, the test uses information that was not available on the simulated date. A study can suffer from both at once — and both inflate historical results.

Is a backtest on current official data automatically contaminated?

Often, yes. Revised macroeconomic series and restated balance sheets make today's version of the past differ from the version that was public on each date. The relevant question is always: was this number, in this version, available on that day?

How does a reader spot the problem in someone else's study?

By looking for the timeline of the inputs: does the study state when each data point became public? Does it use original or revised versions? Does it describe how it avoided decisions made with knowledge of the outcome? The absence of those answers usually says more than any performance curve.

Does the bias only affect quantitative models?

No. Any narrative about the past written with the outcome in hand — "it was obvious that was coming" — commits the verbal version of the same error. The literature calls that cousin hindsight bias.

If an honest test requires dating every piece of information, the next step is understanding where dated research becomes public before any formal seal: what a preprint is and where ours live →

House readings: today's note, in the Daily · the precedents, in the Atlas.

Auditing the timeline of a specific test is the kind of exercise the house conducts in conversation, outside the public article.