Methodology
How Copafever analyses a match
A quantitative model first computes the match probabilities from real data. The AI then writes the read around those numbers. Never the other way round: the AI is not allowed to invent a figure.
Kept up to date with the app's code.
1. The data
Every analysis starts from the official match data, provided by API-Football and refreshed several times a day:
- both teams' form (last 5 games, points, goals scored and conceded);
- the table and season results, split home and away;
- head-to-head record;
- announced injuries and suspensions, and line-ups once known;
- the market consensus (see point 3);
- in the 48 hours before kick-off, what the press says.
2. The model: 81 scorelines computed
Each team gets a strength rating from its league position, adjusted for recent form and its home or away results this season. The gap in strength gives each team's expected goals.
Those expected goals feed a Poisson distribution, the standard model for football scores: we compute the probability of each of the 81 scorelines from 0-0 to 8-8. Adding up the right cells gives each team's win, the draw, over/under 2.5 goals, both teams to score and the most likely score.
3. The market as a safeguard
For the result, the model is combined with the market consensus: we read every operator pricing the match, remove their margin and average them. When at least 5 operators agree, the consensus carries 60% of the final probability; less when there are few.
Why say so? Because it is true, and because the market carries information (late injuries, rotation) no model has on its own. The analysis also shows the gap between our model and the market: when they disagree, that is often where the match gets interesting.
No odds are shown and no betting site is mentioned: Copafever reads a match, it does not sell a bet.
4. The AI's role
Once the numbers are set, Claude (Anthropic) writes the read: match scenario, key factors, key duels, players to watch, impact of absentees. It receives the model's numbers and may not produce others.
The press review uses a web search capped at 3 queries. Betting sites are excluded and every link is checked against the results the search actually returned: an invented link cannot appear.
5. Checked in public
Every analysis written before kick-off is compared with the final score, misses included, on the Accuracy page. To date: 60% correct reads of the result (21 of 35). Picking at random among three outcomes gives about 33%.
Football stays uncertain: the model is regularly wrong, which is why the misses stay on show.
FAQ
Is Copafever a betting site?
No. No odds shown, no betting site, no staking advice. Copafever analyses matches with data.
Where does the data come from?
Official competition data via API-Football (results, tables, squads, injuries), the market consensus and, before the match, the press.
How reliable are the analyses?
60% correct reads of the result over 35 checked matches, match by match on the Accuracy page.