What is being predicted
For each upcoming match we estimate the chance that it is level at the end of regular time (90 minutes plus stoppage time). Extra time and penalty shootouts are ignored: a match that needs them was level after regular time.
Where the data comes from
Fixtures, results and league tables come from football-data.org (European competitions) and API-Football (African competitions). African leagues are covered by a thinner data feed: we hold fewer past matches and no league tables for them, so their estimates are less certain and a match is only predicted when both teams have enough stored results. We store them, record when they were fetched, and never invent a fixture, statistic or percentage. If data is missing, we publish nothing.
The estimate
Three signals are blended:
- Goal model (50% weight). Each team gets an attack and a defence rating from its past results. Recent matches count more (half-life 240 days), and ratings are pulled towards the competition average when a team has few matches. These ratings give expected goals for each side, and a Poisson score model (with a small low-score correction) gives the chance that both sides score the same number of goals.
- Team draw tendency (25% weight). How often each team’s own matches end level, again weighted towards recent games and pulled towards the competition rate.
- Competition draw rate (25% weight). How often matches in that competition end level. This is also our baseline.
Head-to-head results are shown on match pages but are not used by the model: meetings are rare, and squads change.
When enough backtest data exists, the blended number is passed through a simple calibration so that estimates match observed draw rates. The model is a statistical model, not artificial intelligence, and we do not describe it as such.
What gets published
- Both teams need at least 8 past matches, and the competition at least 100.
- The estimated draw chance must be at least 26%.
- At most 8 new predictions are published per day, highest chance first.
- Predictions are made up to 7 days ahead and may be refreshed as new results arrive.
- 60 minutes before kickoff each prediction is locked. A locked prediction cannot be edited or deleted, and the database enforces this.
Signal labels
Mild, moderate and strong describe how far a match’s draw chance sits above its competition’s average (under 2.5, 2.5 to 5, and over 5 percentage points). They are not a measure of how likely we are to be right.
How results are settled
- Correct: level after regular time.
- Incorrect: one team ahead after regular time.
- Void: postponed, cancelled, awarded off the pitch, or no result after 3 days. Void predictions are kept, labelled with the reason, and left out of accuracy.
Accuracy is correct draw calls divided by settled predictions, always shown with the number of predictions.
How we test the model
We replay past seasons week by week, training only on matches played before each week, so the model never sees the future. We compare its probabilities with the plain baseline of using the competition’s draw rate for every match, using Brier score and log loss, and check calibration. The latest run covered 5,218 out-of-sample predictions from 2023-11-02 to 2026-11-14. See the accuracy page for the numbers.
Limits
- Draws are hard to predict. Even a good model is wrong on most individual matches.
- We do not have injury, line-up or odds data, so the model cannot see them.
- Backtests describe the past. They do not promise future accuracy.
- Nothing here is a sure bet, a fixed match or guaranteed profit.