How return, CLV and admission to the rating are calculated. Every number on this page is read from the same settings as the calculation itself, so the text cannot drift away from it.
The statistics include a published pick with a stated event, market, selection and stake. A draft does not count: until a pick is published, it does not exist.
A pick is accepted only before kick-off. The cut-off is taken from the provider data: if the match started ahead of schedule, entries close then. Once published, the pick, the odds and the reasoning cannot be edited — that is what makes it verifiable.
A void (a cancelled or postponed match) counts as neither a win nor a loss: it leaves both the numerator and the denominator.
The odds are set by the server at the moment of publication from the reference line, not on the author’s word. The form cannot submit a value of its own — it is taken from the event’s quotes for the same market and selection.
If there is no reference line for the chosen outcome, the pick is not published at all: with no price to lock there is nothing to verify.
The stake is set as a share of the feed bankroll, so profit is counted in the same money rather than in arbitrary units.
ROI answers the question “how much does each unit staked return” and does not depend on volume. Next to it the table shows bankroll growth — “how much the account grew if you followed the author in full”. These are different questions, and both are shown on purpose.
The difference between the locked odds and the odds at the close of the line. It shows whether an author is systematically ahead of the market, rather than whether they guessed one match right.
CLV is more meaningful than ROI on a short history: the return over thirty picks is mostly noise, whereas a consistently positive CLV cannot be. It is computed only where the closing line is known.
Each window has its own minimum: nobody makes three hundred picks in a week, so requiring them makes no sense.
In addition, average odds of no less than 1.3. The threshold rules out building statistics on near-certain outcomes, where high accuracy says nothing about skill.
A feed that does not reveal its picks after the match stays out of the rating: its results cannot be checked from outside, and the rating is about exactly that.
Rows are ordered not by raw ROI but by ROI trimmed by sample size.
The ROI itself is shown unadjusted in the table — only the position in the list is damped. This keeps +40% over a dozen picks from ranking above +8% over three hundred: the former is nearly always luck, and on a short sample it cannot be told apart from skill. The adjustment works symmetrically: a weak result on a short history is pulled towards zero too.
The platform’s models stand in the shared rating under the same rules: the same admission thresholds, the same calculation, no computational privileges. The only difference is a badge on the row.
They publish the way people do: the server captures the odds, the pick is locked before kick-off and does not change afterwards. Retraining a model does not recompute already settled picks.
Part of the models’ history was published retroactively: the verifier arrived later than the models began computing. Such picks are marked separately and count towards the rating.
The grounds for inclusion are the verifiability of the inputs, not trust: only matches whose computation timestamp is strictly earlier than kick-off are restored, and the odds are taken from the opening line — the earliest price available. Matches played while the verifier was already running are not restored at all.
Picks whose pre-match nature is assumedrather than verified are marked separately: the source does not store the time of calculation. They count as well — in the interface such picks carry a different label.
Changes apply from the date they are published and do not recompute settled picks. Otherwise a track record could be “improved” by editing the rules, whereas it is meant to record what actually happened.