How we count
Every number published here carries its sample size, its window, the status of its data and the version of the method that produced it. The version in force is 1.0, in force since 2026-09-25, and the rules below are what the code actually does (source: our own methodology record). MeasuredLab measures promise against fact.
Data as of 25 September 2026 · source: our own methodology record
What a measurement here means
We record what a forecaster published, we price it independently, and we settle it against the official result. Every number on this site therefore has four things attached to it: the sample size, the window, the status of the data and the version of the method.
- measured — we observed it ourselves;
- stated by source — the source says so and we could not verify it;
- calculated — derived from our own records by a published rule;
- unknown — we did not measure it. Unknown is never written as zero.
version 1.0
What counts as a pick and what is executed
A statement with an explicit price and stake is executed in two series: at the price the author stated and at our own market snapshot. A statement that names a pick without a price is executed at our market snapshot only, and the return at the author's own price is reported as not stated. A correct-score forecast is measured for score accuracy, and a separate derived series settles the 1X2 outcome implied by that score — that derivation is our rule, not the author's claim, so it is reported apart from the author's own market return. An opinion, a conditional or a joke is kept as an observation and never executed. A statement that is not about an event is not counted at all.
version 1.0
What is executed and what is not
A statement is executed only on an exact match of event, market, line and pick, and only if we saw it before the event started. Our timestamp is the moment the material entered our archive: a publication time at the source can be rewritten, ours cannot. A statement without a price is executed at the market price at the moment of the decision; if no independent quote was available, nothing is executed and the reason is published. The same pick repeated across several channels is one observation and one bet per series. Reprocessing the same signal never creates a second bet: the operation key is computed from the content itself and the database rejects a repeat.
version 1.0
ROI
ROI is the net result divided by the total staked, at an equal stake of 10 points per bet. We publish it twice where we can: at the price the author stated, and at our own independent market snapshot taken at the moment of the decision. Void bets are excluded from the denominator and reported separately. The two returns are compared on a common sample, and the share of each sample that falls into that common set is shown next to the number.
version 1.0
Confidence interval
Bets on the same match are not independent, so the 95% interval is produced by a clustered bootstrap over matches with a fixed seed. Where there are too few matches to resample, the interval is reported as not estimated rather than guessed.
version 1.0
The simulated bankroll
Every tracked author starts with 1000 points and stakes 10 per qualifying pick. When the free balance falls below 10 and no bet is open, 1000 points are added and the remainder is kept. Top-ups are counted and drawn separately and are never mixed into the result. No money is involved anywhere on this site: the bankroll is a unit of measurement, not a wallet.
version 1.0
Settlement
An outcome is settled from the official result of the event under the rule published for that market: 1X2 and totals are settled on regulation time only. A match that was abandoned, awarded or never played is void, not a loss. A restatement of the result by the data provider produces a new version of the settlement, not a second bet: the previous version is reversed in the ledger and the change is visible in the corrections log.
version 1.0
Rankings
A ranking shows a window, a minimum sample size and the number for each row. Rows are ordered by that number alone, and the order is computed before any commission is known. An author with no qualifying pick for 30 days keeps the profile with a note and leaves the current window. Nothing in a ranking is for sale.
version 1.0
Reproducibility
Every published number can be recomputed from the journal of bets and outcomes, and the recomputation has to produce the same number — that is a test the build runs against itself. Each metric carries the version of the method that produced it, and a change of method changes the version rather than the past.
version 1.0
The same data, machine readable
One snapshot: these files carry exactly the numbers shown above.