Loading…
Pick a sport. Open the model. Every match gets a projected result, win probabilities, and the reasoning behind them — and past projections are scored against what actually happened.
This page grades the same game projections the dashboard shows — the predicted winner, its stated win probability, and the predicted score — locked to a versioned file while kickoff is still in the future, then settled against official final scores. Wherever market data is stored, the model is benchmarked against the bookmaker favourite on the same games, because "right more often than the market" is a claim that has to be earned, not implied. This measures projection accuracy, not betting returns. A sport with too few settled games shows its live sample count instead of a percentage — nothing here is estimated, backfilled, or illustrative.
Every settled projection, wins and losses alike. Each one was locked to a versioned file while kickoff was still in the future and graded against the official result — never backfilled, never quietly dropped when it went badly. This is a record of accuracy, not of betting returns: we publish how often the projected winner was right, not what a bet on it would have paid.
Every sport on this site runs on its own pipeline, but they all follow the same contract: real data in, a stated method in the middle, and a probability out that we then score against what actually happened. This page explains the method. The live numbers it produces are on the Accuracy page.
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, we say so rather than dress it up.
Sports betting markets are the strongest public forecast that exists — they aggregate every model, insider and sharp bettor in the world. A projection site that pretends otherwise is selling you something. So for the sports where our signals are graded against bookmaker prices, we don't ask you to take the raw model on faith: we refit a calibration on settled results that blends the model's raw probability with the de-vigged market consensus, and keeps whichever mix actually predicts outcomes best under cross-validation.
We are plain about where that lands today: in the current AFL and MLB fits, the market carries most of the weight — the published win probability is closer to a recalibrated market consensus than to the raw simulator, because that is what the settled history supports. The NBA model builds the same humility in by design, anchoring its simulated mean halfway to the bookmaker's line before any probability is computed. When the model earns independent weight on settled data, the refit gives it that weight automatically.
This is the same story the Accuracy page tells with live numbers: its calibration table compares stated confidence against actual win rates, bucket by bucket, on settled results only. A well-calibrated 60% means winning about 60% of the time — that, not a hero win rate, is what we optimise for.
The tier label on a match card is a plain function of the calibrated win probability — nothing more. Two-outcome sports mark a projection HIGH from a 68% win probability and MED from 58%; sports where the draw is live use lower cutoffs (HIGH from 55%) because a 55% favourite in a three-way market is a strong call. NRL projections carry the model's own tier, set at 70% and 55%.
A tier is only worth showing if it means something on settled games. For MLB, a daily check verifies that HIGH projections actually win more often than MED, and MED more often than LOW — if that ordering inverts, the day's publish is blocked. And when tier labels have failed validation — as happened with AFL player-prop confidence tiers — we remove them from the site rather than keep showing a label the data doesn't back.
A methodology is only as credible as the machinery that stops it quietly failing. Four mechanisms run without human discretion:
The output of all four is public: the Accuracy page is the receipts.
Edge Analytics is a decision-support projection tool. It shows you 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'd rather show you "tracking" than a number that means nothing yet.
18+. Bet responsibly. You are solely responsible for any wagering decisions you make. Gambling Help Online: 1800 858 858 · gamblinghelponline.org.au