Each model ingests its sport's raw data, produces a projected result, and converts it into a win probability. The engines differ by sport — and where a model is simple, or where the numbers are not ours at all, we say so rather than dress it up.
Data in. Match results from 2015 to the present, nrl.com fixtures, and a consensus of Australian bookmaker head-to-head odds.
Model. A gradient-boosted classifier for the win probability — with Elo ratings among its input features — plus separate boosted regressors for margin and score. The classifier's output is probability-calibrated (Platt scaling) as part of training.
Probability. The published probability is the model's own. Market consensus is used only as a reference: a projection is flagged VALUE when the model diverges from the de-vigged consensus by 10 points or more.
Data in. AFL Tables player statistics, the Footywire injury list plus manual late-out overrides, venue weather, and sportsbook prop odds.
Model. A projection blend of prior-season and current-season rates with recent-form windows, opponent pace, and tagger-matchup adjustments. Single-leg probabilities come from Poisson or Normal distributions per stat type; correlated multi-leg prices share one simulated game state.
Probability. Every win probability is passed through a calibration fitted on settled signals before any edge is computed — no raw model probability reaches the site.
An important caveat. AFL match-winner probabilities on this site are derived from de-vigged bookmaker odds, not from a model of ours. The graded pick and the market favourite are therefore the same side by construction. We label that market-derived everywhere it appears, and we do not present it as evidence of an AFL model.
Data in. Try-by-try match records since 2023, nrl.com team lists, and try-location data.
Model. Deliberately not a predictive model — it is a matchup tool. Player try rates are weighted 70% last six rounds / 30% full season and set against the opponent's positional defensive record.
Output. Signals are graded A/B/C by fixed threshold rules on rate and opponent rank. They are rankings of historical matchup strength, not win probabilities, and they are not presented as either.
Data in. Statcast pitch-level data, FanGraphs advanced metrics, MLB Stats API schedules, lineups and results, and sportsbook lines.
Model. A 20,000-run game simulation: expected runs from park factors, starting pitcher and bullpen quality, and weather; player props from per-plate-appearance rates and matchup adjustments.
Probability. Raw simulator probabilities are never published. Each market's win probability passes through a calibration fitted on hundreds of settled finals, and the published totals and margins carry fitted bias corrections. Games without a pre-game market are excluded from the slate entirely.
Data in. Official NBA box-score feeds, ESPN injury reports, defence-vs-position splits, and sportsbook player-prop lines.
Model. A full game simulation (10,000 runs per slate by default): possessions and pace, projected minutes from recent rotations and injuries, and per-minute player rates adjusted for matchup and form.
Probability. The simulated mean is deliberately anchored halfway to the bookmaker's line before over/under probabilities are computed, and book odds are de-vigged to fair probabilities. Signals that fail sanity checks — a projection implausibly far from the line, a degenerate simulation, invalid odds — are dropped or downgraded to PASS.
Data in. Understat expected-goals data, player availability from the FPL API, and confirmed line-ups shortly before kickoff. International football fits the same engine on goals.
Model. A Dixon-Coles Poisson model — the standard academic football model — fitted with recency weighting.
Status. Both are marked tracking: no settled rows are graded yet, and no market odds are stored, so there is no baseline to measure against.
Four mechanisms run without human discretion.
Round projections are snapshotted to a versioned, committed file only while every kickoff is still in the future. A snapshot that cannot be proven pre-match is voided and excluded from every statistic on the site.
Outcomes come from official sources — nrl.com final scores, MLB and NBA box-score feeds, Squiggle for AFL — matched back to the locked projection. Losses are stored and shown alongside wins; nothing is backfilled.
When a market's settled results show the model losing its edge there, it is suspended in config and stops being published — as AFL goal and tackle props, AFL game totals and MLB home-run props are today. Suspensions are code, not judgement calls made after the fact.
Before AFL and MLB signals reach the site, an automated check re-scores calibration on settled results. If stated probabilities drift more than 10 points from reality in any bucket, or the model scores worse than a no-skill baseline, the publish fails and nothing ships that day.
Edge Analytics is a decision-support projection tool. It shows what calibrated models expect and how those expectations have scored historically. It is not gambling advice, and no probability on this site is a guarantee of anything: a 70% favourite loses three times in ten, and past model performance does not guarantee future results. Sports without enough settled history show their live sample count instead of a record — we would rather show you tracking than a number that means nothing yet.
This section lives there too — you were sent a direct link to it.