How Market Reads Work

A market read is one row per outcome of a match: the market's probability for that outcome (derived from the best available odds, with the bookmaker's margin stripped out) shown next to the model's own calibrated probability. Where the two diverge, that divergence is displayed transparently — as model output.

There is no selection filter. Every match the model can price produces reads for every outcome of every market it evaluates — favourites and long shots, agreements and disagreements alike. Nothing is cherry-picked, and every read settles publicly against the actual result.

A read is a model view, not a prediction and not a claim that the market is wrong. Divergence is not an advantage.

Decimal odds are the bookmaker's price for an outcome. They encode an implied probability:

Implied Probability = 1 ÷ Decimal Odds

A line of 2.50 implies a 40% chance. Bookmakers add a margin (the overround), so the implied probabilities across all outcomes in a market sum to more than 100%. The market percentage shown on each read is the de-vigged probability — the margin is removed so the outcomes sum to 100% and the number reads as a genuine probability.

Odds are fetched from multiple providers across multiple bookmakers; for each market the best available price is selected, and the read records which source supplied it.

The model estimates each outcome's probability independently of market pricing, using a sport-specific statistical engine:

• Football — a Dixon-Coles Poisson score matrix fit on multi-season results (blended with Elo ratings where fixture history is sparse), accounting for rest days, cross-competition form, and knockout context.
• Tennis — surface-adjusted Elo ratings derived from match-level results on each court type, with recent-load context.
• NBA — a Gaussian efficiency model built on rolling scoring and shooting-efficiency ratings, with playoff-aware variance.
• NFL — a Gaussian efficiency model on rolling scoring, with rest and short-week context.

Model outputs pass through calibration layers fit on historical outcomes, so a displayed 60% is intended to land about 60% of the time over a large sample. Calibration is measured continuously and published on the public track record.

On each read, the de-vigged market probability is the headline forecast. The model appears as a signed lean — the gap, in percentage points, between the model's probability and the market's (▲ model higher, ▼ model lower).

Why is the market the headline? Because we measured it: on our settled-read history, the de-vigged market price has been at least as accurate a forecaster as the model on average. Presenting the model as the headline would overstate it. The lean shows you exactly where — and by how much — the model disagrees, so you can weigh that disagreement yourself.

A lean is a directional model view, not a proven advantage. Sports markets are hard to beat, and we do not claim to beat them.

Every read is settled against the official result and kept — hits and misses alike. Football 1X2, totals, and BTTS settle on the 90-minute score; walkovers, retirements, and abandoned matches are voided, never silently dropped.

Quality is measured the way forecasters are measured: calibration (does 60% happen 60% of the time?) and Brier score, not highlight reels. Individual outcomes are subject to variance — a read that does not land is not a model failure; it is the expected cost of forecasting under uncertainty.