From real data to a probability

Every day we collect results, team and player statistics and the Bet365 odds for the next 7 days across football, basketball, handball and tennis. Our engine turns that history into a probability for each market, and then compares it with the price the market is offering. Three steps: real data → statistical model → comparison with the market.

Poisson / Dixon-Coles with attack and defence

For football we estimate the goals each team is expected to score, combining that team's attacking strength with the opponent's defensive weakness. A Poisson model (with the Dixon-Coles correction for low-scoring games) turns those expected goals into the full grid of outcomes: 1X2 probabilities, over/under lines, both-teams-to-score and the most likely correct score. We do not use raw averages — every number is adjusted for the quality of the opponent faced.

Opponent-adjusted Elo

An Elo rating measures the level of each team or player and updates after every match. Ours is adjusted for the quality of the opponent: beating a strong side is worth more than beating a weak one. This rating anchors the base strength of the model in every sport.

Basketball and handball

Here we combine opponent-adjusted Elo with each team's pace and recent points for and against, to estimate the total points/goals of the game against the market line.

Elo by surface

A player can be much stronger on one surface than another, so we keep separate Elo ratings for clay, hard, grass and carpet, adjusted for the opponent. From there we estimate the favourite and, when the sample is reliable, the total games of the match.

The live radar and comeback probability

Beyond pre-match predictions, our live radar watches the matches being played right now and flags the two moments where the pre-match script has broken: a favourite in trouble (the clear market favourite trailing or level when it should be ahead) and goals that aren't coming (a match that pointed to plenty of goals still stuck on few or none as time runs out).

For each of those matches we compute a comeback probability —also known as comeback or turnaround probability— the real chance of the scenario straightening out, conditioned on the current score and the time left. We use a specific model per sport: in football, a goals model (Poisson conditioned on the score and the minute), which beats the naive baseline in our checks (Brier score 0.242 vs 0.25 for the trivial forecast); in tennis, a Markov chain over sets and games (best of 3 or best of 5); and in basketball and handball, the points difference and remaining time, calibrating forward as we gather more in-play samples.

The result is shown in four colour bands so it reads at a glance: very low (red, up to 39%), low (amber, 40% to 54%), medium-high (blue, 55% to 64%) and high (green, 65% and above). Crucially, this is a data point to analyse — it is neither a pick nor a betting odds figure. It tells you how likely the turnaround is according to the model; whether that's worth anything against the live price, and whether to bet at all, is always your call.

Every live alert has a closed, auditable life cycle. It is born when the radar detects the scenario (a favourite in trouble or goals not arriving) and computes its comeback probability; at that moment it is frozen exactly as issued, with its score, its probability and the live odds of that instant. While the match goes on, the probability is recomputed every minute and the alert can be promoted as it crosses into a higher band: each subscriber is alerted exactly once, at the moment the probability crosses the filter they configured — never two alerts for the same match and market. When the match ends, the alert is settled with its real outcome ( if the match got back on script, if not) and joins the day's log, where we publish the stats by band and by filter — number of alerts, average odds, units and yield — exactly as they come out, bad days included. That log is the proof of what the alert system did, not a promise of what it will do.

Comparison with the market

We anchor our probability to the real 1X2 from Bet365 and show, market by market, where we agree with the price and where we differ. The odds movement (open → close) is also shown, because how the market moves is one of the best judges of a prediction.

Honesty by default

The betting market is very efficient and no model wins every time. That is why we do not sell "sure" picks: we show our probability next to the market's, match by match — even when they agree and there is no edge — and we publish the model's real performance on already-settled value bets. The data is ours to give you; the decision is always yours. See the track record →