There is no single right way to predict a match. Some look at recent form, some at the whole season. Some trust the goals that were scored, others say "goals are luck, look at the shots". Each view is right in some matches.
In LigLab, each of those views is a separate bot. Seven bots look at the same match from seven different places, each writes its own prediction, and every one of their records is kept in the open. This article explains the seven bots and the differences between them.
Common ground: team strength and score probabilities
Most of the bots start from the same basic idea: every team has an attacking strength and a defensive strength. A strong attack scores a lot, a strong defence concedes little. When the strengths of two teams and the home advantage are put together, you get how many goals each team is expected to score in that match.
From the expected goals, the probability of every scoreline is calculated: how likely 1–0 is, how likely 2–1, how likely 0–0. The sum of those score probabilities gives the three outcomes: home win, draw, away win.
What separates the bots is what they look at when measuring team strength.
Four measurements: goals, shots, form, rating
Goal Counter: looks only at goals
The plainest bot. It works out each team's strength from the goals it scores and concedes, giving older matches less and less weight. Its slogan says it all: "Looks only at goals. The rest is story."
Its strength is its simplicity; its weakness comes from the same place. In a low-scoring game, a few lucky goals can make a team look stronger than it is.
Sharpshooter: goals lie, shots on target don't
It does the same calculation as Goal Counter, but looks largely at shots on target instead of goals. A team with lots of shots and few goals is placed where it "deserves" to be; lucky goals are filtered out.
Shots happen far more often than goals, so they give earlier and sturdier information about a team's real strength.
Form Watcher: what do the last weeks say?
The same strength calculation, but with a short memory. It gives much more weight to the matches of the last few weeks. A team recovering under a new manager, or sinking under a string of injuries, is noticed before the other bots notice it.
It pays for that too: it can take a few good matches too seriously.
Rating Keeper: every match is a chess move
It works like chess ratings. After every match a team's rating goes up or down depending on the score and the strength of the opponent; home advantage is factored in. Beating a strong opponent earns a lot, beating a weak one earns little.
It doesn't go into goals and shots; it only asks "who beat whom, and by how much", and it never forgets a match.
Consensus: weighs them all and combines them
It exists to be the most consistent of the seven. Instead of tying itself to a single view, it combines two, weighting them: about 70 percent goes to the team strength derived from goals scored and shots on target, and 30 percent to the rating updated match by match.
On top of that it adds the conditions of the match: rest time, big matches, early-season uncertainty, matches played behind closed doors. When the squads are announced on match day, it updates its probabilities for the eleven who actually take the field.
What we call "LigLab's prediction" on the site and in the app is Consensus's prediction. The figures in our locked tests are its figures too; its current record is on the Record page.
Two different goals: Points Hunter and Upset Seeker
The last two bots don't take a new measurement. They take Consensus's probabilities, but do something else with them.
Points Hunter: looks for the score that earns the most points
In a prediction game the aim isn't only to know who will win, but to collect points. Points Hunter works out the expected points of every scoreline and picks the highest.
In close matches this pushes it towards draws such as 1–1. Its result accuracy drops a little, but the points it collects go up. If you're chasing points in a friends' league, this is the bot to follow.
Upset Seeker: doesn't trust the favourite easily
An experimental bot. It uses the same probabilities, but puts extra weight on the results that aren't the favourite: draws and wins for the weaker side.
Its record is deliberately kept open: it shows the price of looking for surprises. When it lands, people talk about it a lot; in the long run it trails those who trust the favourite.
The seven bots at a glance
| Bot | What it looks at | When it stands out |
|---|---|---|
| Consensus | Weighted blend of strength and rating, match conditions | Most consistent in the long run |
| Goal Counter | Goals scored and conceded | In plain, steady teams |
| Sharpshooter | Shots on target | In teams that are running lucky or unlucky |
| Form Watcher | The last few weeks | In teams whose form is changing fast |
| Rating Keeper | Results and the strength of the opponent | In matches with a clear balance of strength |
| Points Hunter | The score that earns the most points | When collecting points in a prediction game |
| Upset Seeker | Results other than the favourite | In weeks of surprises |
Why not just one bot?
There are three reasons.
First, no single view is right all the time. What Sharpshooter sees, Goal Counter doesn't; the change Form Watcher catches, Rating Keeper notices late. When the bots split on the same match, that is information too: if they all point to the same team the match is clear, if they split it is uncertain.
Second, combining beats any single one. That is why Consensus exists: the errors of different measurements partly cancel one another out.
Third, comparing keeps us honest. When the records of seven bots stand side by side, you can see which view really works. A single "magic model" claim never goes through that test.
Your turn
The bots' predictions for this week are on the league pages: Süper Lig, Premier League, La Liga, Serie A, Bundesliga. In the app you can enter your own prediction and compete on the same table as the seven bots.
How the models are tested and which ideas failed the test is on the Method page.
In short
- Seven bots look at the same match from seven different places: goals, shots on target, form, rating and a blend of them.
- Consensus combines the measurements by weight; "LigLab's prediction" is its prediction.
- Points Hunter and Upset Seeker take no new measurement; they pursue another goal with the same probabilities.
- A match where the bots disagree is an uncertain match.
Sources
- Dixon, M. J. and Coles, S. G. (1997). Modelling association football scores and inefficiencies in the football betting market. Journal of the Royal Statistical Society: Series C, 46(2).
- Maher, M. J. (1982). Modelling association football scores. Statistica Neerlandica, 36(3).
- LigLab: Method and How we work.