How Tips8ng predictions work

In short

For each upcoming match in the competitions we cover, a statistical model estimates the probability of a home win, a draw and an away win. Those probabilities are a forecast, not a result and not a guarantee. After the match, every published prediction is graded against what happened, and the record is public, including the misses.

What goes in

Past results and match statistics from a commercial football-data provider. For each team, the model uses only information available before kickoff:

• a long-run strength rating (Elo), updated after every match;
• recent attacking and defensive performance, adjusted for the strength of the opponents faced;
• recent form overall, and at home or away;
• how strong the recent opponents were (strength of schedule);
• recent shots on target, possession and expected goals (xG);
• rest days since the last match, and home advantage.

Player availability (injuries) was tested as an input and rejected: it didn’t improve the forecasts. The model doesn’t read news, line-ups or betting prices.

The models

Two main kinds of model are used, and each competition gets the one that forecast better in a walk-forward test on its own past seasons (train on earlier matches, test on later ones):

• a Poisson goals model, which estimates each side’s expected goals and turns them into a grid of scoreline probabilities. As of the review date it serves La Liga, Serie A, the Bundesliga, Ligue 1, AFCON, the World Cup and CONMEBOL competitions;
• a Random Forest classifier, which predicts home, draw or away directly. It serves the other predicted competitions, including the Premier League and the PSL.

One exception: the Carabao Cup uses a simpler cup model built on each club’s division, because form-based models can’t compare clubs from different divisions. It passed the same kind of walk-forward test, but it is less sure of its calls, and the competition page says so.

Some competitions are shown with fixtures and results only, without predictions: either no model passed the test there, or the competition hasn’t been evaluated yet. If a model is unavailable for a match, a simpler baseline estimate is shown instead and is labelled “baseline estimate”; baseline estimates are left out of the public record.

What comes out

• Home / draw / away probabilities: the core forecast.
• A confidence level from how far the top outcome is ahead of the others. When it is low, the card says “No strong lean” rather than naming a winner.
• A projected scoreline, chosen so it agrees with the called outcome; a narrow call is shown as a narrow score.
• Both teams to score and over 2.5 goals are shown as each competition’s long-run rate, not a per-match model output: per-match versions forecast worse than a constant rate when we tested them.

How we check it

Every published prediction is settled on the 90-minute result and scored with proper scoring rules (the Brier score and log loss) and calibration: when we say 60%, does it happen about 60% of the time? It is compared with two simple baselines (an even three-way split, and each outcome’s historical frequency) and, where we captured them, with bookmakers’ prices with the margin removed. The model record shows all of it, and the prediction results page shows each match.

Honest summary as of the review date: the models beat the simple baselines modestly, and the bookmakers’ margin-free prices have been the more accurate forecast. Check the record for the current numbers.

Limits

• Football is noisy: a 35% outcome happens about one time in three. One match proves little either way.
• Early in a season, and in cups between teams that rarely meet, there is less data and the forecasts are weaker.
• The model knows nothing about late team news, injuries or tactics announced before kickoff.
• Nothing here is betting advice or a promise of profit. If you bet, read responsible play (18+).

Written content and AI

Match previews and analysis text on Tips8ng may be drafted by an AI language model from the model’s numbers, with rules that reject claims the data doesn’t support. Blog articles go through human editorial review before they are published, and every social media post is approved by a person.