Glossary
Updated 2026-09-27
Plain-language definitions of the terms used on AI Ball's match pages.
- Win/draw/loss probability
- The model's estimate, before kick-off, of each of the three results. The three numbers add up to 100%.
- Confidence
- How far apart the three probabilities are and how many of the model's views agree, mapped through a table fitted on past matches. It is not the chance of being right.
- Market view
- Probabilities read from the prices that dozens of sources quote for a match, after each source's margin is removed.
- Margin removal (Shin's method)
- A published way (Shin, 1993) to strip the built-in margin out of quoted prices so that the implied probabilities add up to one.
- Expected goals (on AI Ball)
- A pre-match estimate of how many goals each side will score, derived from the goal-line market and the sides' relative strength. It is not shot-based xG.
- xG (shot-based expected goals)
- A post-match measure that adds up the scoring chance of every shot. AI Ball does not use shot data, so it does not publish xG.
- Poisson model
- A standard way to model counts of rare events; in football, the number of goals a side scores (Maher, 1982).
- Dixon–Coles model
- An adjustment to the Poisson scoreline model (Dixon & Coles, 1997) that corrects the under-count of 0-0, 1-0, 0-1 and 1-1.
- Elo-style rating
- A strength rating that moves up or down after each result, by more when the result was less expected.
- Most likely scores
- The five scorelines with the highest probability in the model's score matrix.
- Analysis mode
- One of six weightings of the same inputs. Balanced is the default and the only mode used on the open record.
- Risk level
- A low / medium / high / critical label built mainly from low confidence, disagreement between the model and the market favourite, and league history.
- Locked pre-match view
- The model's numbers as archived 45–120 minutes before kick-off. They cannot be edited or deleted, and they are what the open record is marked against.
- Always-the-favourite baseline
- What the hit rate would be on the same matches if you always sided with the side the market rated most likely. It sits next to the model's own column on the open record so the two can be compared band by band.
- Calibration
- Whether stated probabilities match reality: of all the times a model says 70%, does it happen about 70% of the time?
- Brier score
- The mean squared gap between stated probabilities and what happened; lower is better (Brier, 1950).