Football predictions: understanding the model probabilities
HEXAGONE estimates 1X2 probabilities using goals scored and conceded, match venue and league averages. A projection describes a distribution of possible outcomes; it promises neither a result nor a profit.
What goes into the calculation
The model uses the season’s complete result matrix. It separates home and away performance, then combines both clubs’ attacking and defensive rates. A prior equivalent to five matches reduces the effect of small samples. Ligue 2 retains v1 with league averages. In Ligue 1, v2 uses 2024/25 and 2025/26 historical home/away profiles, shrunk with eight historical league-prior matches; match weights halve every 365 days. Retrospective comparisons and coverage are published in the Lab. A projection requires at least 30 league matches and three matches per team. These thresholds permit a calculation; they do not make an early-season sample robust.
From mean goals to outcome probabilities
Two independent Poisson distributions produce a score grid. Cells where the home team scores more contribute to 1; equal scores contribute to X; the remaining cells contribute to 2. The same grid provides totals, BTTS and double chances. These events overlap: their probabilities must not be multiplied as independent events. Model mean goals are not shot-based expected goals (xG).
Reading theoretical odds
No-margin decimal odds equal 1 divided by probability. A 50% probability therefore corresponds to 2.00. This calculation is not a bookmaker quote. The comparison tool accepts your three 1X2 odds and removes their implied margin proportionally. A disagreement with the model may reflect its limitations; it does not establish a profitable opportunity.
How quality is evaluated
On the 2022/23 and 2023/24 archives, each date is predicted using strictly earlier dates after a 60-match warmup. Every match sharing a date is evaluated before any of those scores are ingested. The Lab shows Brier score, log loss, a league-frequency baseline and calibration-bin sample sizes. This is retrospective evaluation, not an independent test on future fixtures.
What the model does not know
Injuries, line-ups, fatigue, cards, shots and minutes are not included. Projections cover eligible future fixtures within 28 days, later than the results cutoff. A fixture already present in the result matrix is excluded. Always read the source date and sample sizes beside each projection.
Are BTTS frequencies predictions?
No. They describe the share of observed matches in which both teams scored. Next-match probabilities are computed separately by the model and depend on the opponent.
Where do Ligue 1 and Ligue 2 statistics come from?
Current results use Wikipedia score matrices reconciled with standings. Dated fixtures and squads have separate source revisions. Historical results come from OpenFootball. Source links and licences are retained.
Are the displayed odds bookmaker quotes?
No. Theoretical odds equal 1 divided by model probability, without margin. The comparison tool receives only the odds you enter; no bookmaker feed is connected.
Why do tables show N or 4/7?
A percentage alone hides sample size. N is the number of matches, and 4/7 means four occurrences in seven games. Missing values are not replaced with zeros.
How are the latest five results selected?
By verified date, after matching the fixture and score against the matrix. With partial dated coverage, these are the latest five available results, not necessarily the latest five matches played.
Can I read the data in French?
Yes. Every page and guide has a French and English version, connected by the language switch. Exports use stable identifiers shared across both languages.