EV (Expected Value) Metric
Expected value (EV) compares the model’s estimated hit probability with the available odds. It is separate from the predicted match direction. A favourite may be likely to win while its win odds are so short that they do not compensate for losses when it fails. A correct single-match direction therefore does not establish long-term value at those odds.
Favourite Win Probability Probability
The model’s estimate of the probability that the stronger team wins within the regular 90 minutes, calculated from both teams’ historical data and live market movement. It is independent of the bookmaker’s quoted odds and is a core measure of match value.
Upset / Draw Probability Probability
The combined probability that the underdog either draws or wins. When this value is materially higher than the market expectation, the underdog or handicap side may offer strong value.
Confidence / Star Rating Rating
A measure of how strongly the available data supports the model’s conclusion. One star indicates a standard recommendation, two stars indicate higher value, and three stars indicate a close match with historically strong upset or favourite-win patterns.
Model Recommendation Output
The site shows match direction and odds value separately. When the model leans toward a favourite, it names the team. If the current odds are too short and EV is not positive, it marks the odds as low value. Directional accuracy and long-term value at those odds should be assessed separately.
Direction Result (Handicap Market) Concept
A “correct direction” in the historical results does not simply mean predicting the match winner. It refers to the recommended side of the football handicap market. For example, if the favourite has only a 51.8% win probability and is unlikely to cover a deep handicap such as -2, the model may favour the underdog on the handicap. A 0–0 result can therefore count as a correct direction even though the favourite did not win.
Market-Implied Probability and De-vigging Market
The reciprocal of decimal odds approximates the market’s raw probability, but probabilities across a market usually add up to more than 100% because of the bookmaker margin. The model normalises away that margin before comparing bookmakers and movements from opening to closing prices.
Expected Goals (Model Mean) Metric
The goals model estimates an average goal count for each team across many comparable matches. This is a pre-match model mean—not traditional shot-quality xG—and not a claim that the match must finish on that exact integer. The two means are then used to construct total-goals and exact-score probabilities.
Poisson Goals Model Model
The Poisson distribution is commonly used for low-frequency count events such as goals in a fixed match duration. Our model uses LightGBM with a Poisson objective to learn the home and away scoring means from recent team and pre-match market features, rather than applying a simple league average.
Dixon–Coles Low-Score Adjustment Adjustment
Independent Poisson estimates can miss dependence between home and away goals. Dixon–Coles specifically adjusts joint probabilities for low scores such as 0–0, 1–0, 0–1 and 1–1 so the score distribution better reflects football data. It is a statistical correction, not a predetermined result.
Probability Calibration and Out-of-Time Validation Validation
Calibration aims to make events predicted at 40% occur roughly 40% of the time over a large sample. Calibrators and recommendation thresholds are fitted only on rolling out-of-fold predictions from earlier seasons, then checked on a later holdout season to reduce future-data leakage.
Total Goals: 0–6+ Classes Output
The total-goals market is split into seven mutually exclusive classes: 0, 1, 2, 3, 4, 5 and 6+, where 6+ means six or more goals. The page shows the two highest-probability classes. All seven probabilities total approximately 100%, subject to rounding.
Top 2 Candidates and Combined Confidence Confidence
Top 2 means the two candidates with the highest model probabilities; it does not mean each has a 50% chance. For total goals, their probabilities are added into a combined confidence score and compared with a league-specific threshold selected on historical development data. Below the threshold, the page says “Skip” while retaining the candidates as reference.
Exact-Score Probability Matrix Probability
Home and away goal counts form a two-dimensional matrix in which every cell is an exact score and its raw probability. Exact scores are inherently harder to predict than match results or total goals, so individual score probabilities are normally modest; that does not by itself indicate a calculation error.
Deterministic Score Re-ranking Integration
The final two league scorelines are not randomly generated by generative AI. A fixed formula reweights the mathematical model’s top four scores using the result direction, the seven-class total-goals distribution and the official handicap relationship. Identical inputs always produce identical outputs. The World Cup section uses a separate workflow.