Naija Score How we predict football matches
Our model turns past results into a probability for each match. This page says what it is, how it is tested, what the numbers mean and what they do not.
What the model is
The model is called naija-poisson-v1. It is a Poisson goals model: for every team it estimates how many goals it tends to score and concede, adds a home-advantage figure for each league, and gives more weight to recent matches than to old ones. From the expected goals of both sides it works out the probability of a home win, a draw and an away win.
It is a statistical estimate, not inside information. It knows nothing about injuries, lineups or motivation, and it does not read bookmakers' odds.
How it is trained without using the future
To predict a match, the model is trained only on matches played before that match. Nothing from the match itself, or from later dates, ever reaches it. This is what stops a backtest from flattering itself with results that were not yet known. The settings of the model were tuned on the 2022 season only, then left fixed for the test below.
What the pick is
When the model is confident enough in one result (at least 60%, and both teams have match history), the pick is that single result: home win, draw or away win. Otherwise the pick is a double chance that drops the least likely result. The percentage shown is the probability of the pick: for a double chance, the sum of the two results it covers.
Next to it we show the fair odds: 1 divided by the probability, to two decimals. A 50% pick has fair odds of 2.00. If a bookmaker offers more than the fair odds, the price is generous compared with our estimate; if less, it is not. We do not store bookmakers' odds, so we cannot tell you where the value is. Compare them yourself.
The backtest
A backtest runs the model over matches that have already been played, as if it had predicted them one by one, each time using only the matches before it. Ours covers 2,576 matches from 2023–24.
| Pick rule from the model (naija-poisson-v1) | 77.8% of picks correct |
|---|---|
| Baseline: always “home or draw” | 72.9% correct |
| Single-result picks only | 73.2% |
| Double-chance picks only | 78.9% |
| Matches in the test | n = 2,576 |
So on past matches the model did a little better than the simple rule of always picking home or draw. Two cautions come with that. First, a double chance covers two of three results, so it is right more often by design. Second, in the Nigerian league (NPFL) the model did not improve on simply picking the home side, so we do not claim an edge there.
Accuracy is not profit
A pick that is right 78% of the time can still lose money if its odds are too low. A double chance usually pays little, so being right often is not the same as winning money. Nothing on this site tells you that a pick has value, and none of it is advice to bet. A 60% pick still loses about 4 times in 10.
What you see on the predictions page
The 40 predictions on the predictions page today are not from the model above. They are the probabilities from our data provider, API-Football, for matches already played, and they were generated after those matches. They are a backtest, not a record of picks published before kick-off, and as single-result picks they were right in 16 of 40 matches. For comparison, always picking “home or draw” on the same 39 matches would have been right 61.5% of the time (24 of 39).
When the live record starts
The live record starts when live match data is switched on and picks are published before kick-off. We will keep it separate from the backtest, and we will call it a live record only after at least 100 of those picks have been settled. Until then, every percentage on the site is labelled as a backtest and shown with the number of picks behind it.
Predictions are estimates of probability. No result is guaranteed. 18+ only. Bet responsibly. Read how we score betting sites and what we do and do not do.