Methodology

Ayght Sports publishes a win probability for each side of every event we cover, and keeps a record of how that number moved. This page explains where those numbers come from, how we check whether they are any good, and what we are not yet claiming.

---

What a probability means

When we put a team at 62 percent, we are saying that across many matchups that look like this one, that team wins about 62 times in 100 and loses the other 38.

That is a different claim from a prediction. A prediction says what will happen. A probability says how often something like it happens. A 62 percent side losing is not the number being wrong — it is the 38 percent arriving, which it should, roughly two times in five.

So the way to judge these numbers is not whether the higher one won. It is whether, across hundreds of events, the things we called 62 percent happened about 62 percent of the time. That property is called calibration, and it is the standard we hold ourselves to.

How a probability is built

Every estimate starts from the record — what teams and participants have actually done, rather than what anyone thinks of them.

• Scoring rates, and how they hold up against different quality of opposition rather than in isolation.
• Point differential across a season, which describes how games were won and lost more honestly than a win-loss record does.
• Home and away splits, because most teams are measurably different in each.
• Recent form, weighted so the last month counts for more than the first.
• Rest, travel and schedule density.

Those inputs feed Poisson-family scoring models and point-differential regression. Rather than producing a single answer, they produce a distribution of plausible outcomes. The probability you see is the share of that distribution falling each way. Championship and season-long markets are simulated across the remaining schedule, then normalised so the field sums to 100 percent.

Estimates are generated independently, on our own schedule. An event can carry a number as soon as it appears on a calendar, often well before any market exists for it.

Fair odds, and what that means

Where we show odds, they are a direct conversion of our probability with no margin added. A bookmaker’s price includes their margin, which is why the two sides of a market there sum to more than 100 percent. Ours sum to exactly 100.

One visible consequence: a two-way market always shows symmetric odds — -253 and +253, never -140 and +120. Both sides come from the same probability, so the magnitudes match and only the signs differ. Asymmetric lines are a product of margin, not of the underlying estimate.

Our pricing method changed on 26 August 2026, when a residual margin was removed. Charts that reach back before that date carry a note saying so. We have not rewritten those earlier values: the original probabilities behind them were not retained at chart frequency, so any correction would be a reconstruction presented as the original record. Labelling it is the accurate option.

What we currently cover

Updated 23 minutes ago

How numbers move, and why

A probability published a month before an event is not meant to be the last word. It updates as the inputs change — form shifts, availability changes, the schedule fills in.

Every event page shows the full recorded history of its number, not just where it stands today. Alongside it we plot a ten-day trailing average, which smooths day-to-day noise and makes the underlying direction easier to see. The rising or falling label is measured against that average rather than against the opening figure, because a number can sit above where it opened while clearly easing.

Where a series begins part-way through, it is because that is when we first published a number for that event. We do not backfill estimates for periods in which none existed.

How we check whether the numbers are any good

It is easy to look accurate in hindsight. A model can be adjusted after the fact, a backtest can be run until it flatters, and a good week can be presented as a track record. We have tried to build so that none of that is possible.

• Predictions are preserved before the event starts. Every probability is snapshotted at the moment it is published. Those snapshots are not editable afterwards.
• Scoring happens after the outcome. Performance is measured against what actually occurred, using proper scoring rules that reward calibration rather than confident guessing.
• Evidence is prospective, not reconstructed. We evaluate on predictions that existed before the result, never on a backtest assembled afterwards.
• New models must earn publication. A revised model runs as a shadow forecast, recorded but not shown, until it has accumulated enough prospective evidence to justify replacing what is live.

We intend to publish the resulting calibration record — how often outcomes we called at a given probability actually occurred, across a meaningful sample — and to keep publishing it whether it flatters us or not. Until that record covers enough settled events to mean something, we would rather say so than show a figure built on too small a sample.

What we are not claiming

This site is in private beta, and several parts of it are still being validated. Being specific about that is more useful than a general disclaimer.

• We are not claiming proven market-beating performance. That is a question the calibration record is designed to answer, not one it has answered yet.
• Championship models for some leagues are running as shadow forecasts and are not yet published.
• Early-season and preseason estimates rest on less current-season data and carry more uncertainty. Model confidence is shown on each event page for that reason.
• Calibration has not yet been established separately for every sport and every forecast horizon.
• Nothing here is a betting recommendation, and no result is guaranteed.

What we do not publish

We do not publish selections, picks, tips, or recommendations. We do not rate one side above another, flag a number as good value, or tell you what to do with any of it.

Both sides of every event are presented identically — same fields, same layout, same emphasis. Where a page has a favourite, that is arithmetic, not endorsement.

The number is the product. The judgement is yours.

Where the data comes from

Probabilities are our own, produced by Ayght Intelligence, our modelling system, which runs separately from everything else we operate. Schedules, results, records and availability come from licensed sports data providers.

Where bookmaker prices appear anywhere on the site, they are kept plainly separate from our own estimates and never blended into them.

Starters, lineups and availability are shown as matchup context. Where they are not inputs to the model, the page says so explicitly rather than letting proximity imply otherwise.

Nothing on this site is official league data or statistics. We are not affiliated with, endorsed by, or sponsored by any league, club, or governing body. Team and participant names appear for identification only.

Methodology versions

Models change as inputs are added. When they do, the numbers change with them, and a figure produced under one version is not directly comparable with one produced under another.

Anything we publish as a calibration figure carries the methodology version and date it was produced under, so a later improvement reads as a documented change rather than an inconsistency. One current example: starting pitchers are not yet an input to the MLB model, which is why three games between the same two teams on different dates can carry the same probability. That is a known limitation with work in progress, not a fault in the arithmetic.

Corrections

If something here looks wrong, we would rather hear it than not. Tell us what you are looking at and what appears incorrect, and we will check it. Write to info@ayghtsports.com.

---

Ayght Sports publishes probability estimates for informational purposes. Nothing here is a recommendation, and nothing here is a prediction of any individual outcome.