How we predict draws: the model, step by step

Every draw chance on DrawMeter comes out of the same statistical model, run on the same data, with no one adjusting the numbers by hand. This page explains it step by step, in plain words, along with the things it cannot know. Nothing here is secret, and none of it makes a draw certain.

Step 1: an attack and a defence rating for every team

The model is a Poisson goals model of the kind described by Mark Dixon and Stuart Coles. Each team gets two numbers: an attack rating, how many goals it tends to score against an average opponent, and a defence rating, how many it tends to let in. Each competition also gets its own scoring level and its own home advantage, because a goal is easier to come by in some leagues than in others, and playing at home is worth more in some places.

The ratings are fitted on finished matches from roughly the last three seasons. Recent matches count for more: a result loses half its weight after about eight months, so last month's form matters far more than a match from two years ago, without throwing the old one away. The UEFA and CAF club competitions link teams from different countries, which is how a Ghanaian side and a Moroccan one, or an English side and a German one, end up on the same scale. Every domestic league keeps its own scoring level and home advantage, while the club cups of one confederation share a single pair.

The model reads scores and nothing else. It does not use expected goals data, Elo ratings, injuries, line-ups or news.

Step 2: expected goals and the score grid

For a match, the ratings give two figures: the goals the home side is expected to score and the goals the away side is expected to score. From those two figures the model works out the probability of every scoreline, 0-0, 1-0, 0-1, 1-1, 2-0 and onwards, a grid of small chances that together add up to 100%.

Everything else follows from that grid. Add up the cells where the home side scores more and you have the home win chance. Add up the cells where the away side scores more and you have the away win. The cells on the diagonal, where both sides score the same, give the draw.

Step 3: the correction for low scores

Treating each side's goals as an independent count is a good approximation for most scorelines, but it fits football worst at the very bottom of the grid: 0-0, 1-0, 0-1 and 1-1. The Dixon and Coles correction adjusts exactly those four scores. For draws this is the step that matters most, because two of those four scores are draws, and together they hold most of a typical match's draw chance.

Step 4: why some matches draw more than others

With the grid in hand, it is easy to see what pushes the draw chance up or down. Take a simple illustration with plain goal counts, before the correction and before any bookmaker blend:

• Two sides each expected to score 1.2 goals: the draw comes out at about 28%.
• Two sides each expected to score 1.5 goals: about 24%. More goals mean more ways for the match to tip one way.
• A side expected to score 2.2 against one expected to score 0.6: about 17%. A clear favourite leaves little room for a level score.

So the highest draw chances belong to evenly matched sides in matches with modest goal expectations. That is why the top of our draw predictions table and the bold eight on the pools coupon tend to be tight games, and rarely a top side against a struggling one.

Step 5: the blend with the bookmaker's price

Where the bookmaker's odds for a match are in our data, the model's match result chances are blended half and half with the chances implied by those odds. First the margin is removed: the bookmaker's three prices imply chances that add up to more than 100%, and they are brought back to exactly 100%. Then each of home, draw and away is set halfway between the model's figure and the bookmaker's.

The odds come from the recommended bookmaker's own feed. Not every match has them: smaller leagues often open their markets only a day or two before kick-off, and matches weeks away may have none yet. Without odds, the draw chance is the model's alone. Correct score chances, such as the 1-1 figures on score draw predictions, are always the model's own numbers and are never blended.

Some pools coupon matches fall outside the competitions the model rates, such as FA Cup ties and internationals between national teams; in a cup week they can be most of the coupon. Those rows are marked "from the odds": their chances are the bookmaker's prices with the margin removed, not a model forecast.

Fair odds and what a probability means

Fair odds are 1 divided by the probability. A 30% draw has fair odds of 3.33, a 25% draw 4.00. They are the price at which a bet would neither win nor lose money over a very long run, if our probability were exactly right. A bookmaker's prices carry a margin, so they are usually a little shorter than fair odds would be.

A probability is a statement about many matches, not one. If we rate ten matches at 30% each, the model expects about three of them to draw; it would not be surprised by one or by five, which between them cover roughly nine cases in ten. None of the ten drawing happens about three times in a hundred. The same logic explains the pools coupon: with eight matches at 30%, two or three draws is the typical result, and all eight together is about a one in 15,000 event.

How the eight likeliest coupon draws are chosen

On the pools coupon, every match with a price gets a draw rank, from the highest draw chance down. The eight highest are printed in bold. There is no other filter, no human choice and no attempt to balance leagues. The ranking and the bold eight are frozen at the week's first kick-off and checked against the final results, Pools Panel verdicts included, on the track record page. How the coupon itself works is covered in how football pools work.

What the model cannot see

Every model is a simplification, and this one has clear blind spots:

• Team news. An injured striker, a suspended centre-back or a goalkeeper returning from injury changes a match. The model knows none of it.
• Motivation. A side that only needs a point, a dead rubber at the end of the season, a manager resting players before a cup tie: the ratings cannot tell.
• Conditions. Heavy rain, a frozen or poor pitch, a long journey or a third match in a week can all push a game towards a cagey draw or away from one.
• Small samples. Early in a season, or for a newly promoted side, the ratings lean on matches from a different division or a different squad, so they adjust slowly to a team that has really changed.
• Clubs new to the data. Teams with few matches in our data get rough ratings, and teams outside the covered competitions get none.

These are reasons to read team news alongside the numbers, not reasons to ignore them.

Data and refresh times

Fixtures, results and odds are refreshed four times a day, at about 02:00, 09:00, 14:00 and 20:40 UTC, which is 03:00, 10:00, 15:00 and 21:40 in Lagos and an hour earlier in Accra. Tables update on their own after each refresh. Bookmaker odds come from Megapari's own feed; the rest of the data comes from these sources:

• scoresults.com: the weekly UK pools coupon, its match numbers and Pools Panel results.
• openfootball: fixtures and results (public domain).
• Wikipedia contributors: UEFA and CAF competitions and some league tables (CC BY-SA 4.0).

A last word on certainty

The model is built to give honest probabilities, and honest probabilities for draws are modest. A well-made draw chance still misses most of the time. Use the numbers to compare matches and to understand a coupon, never as a promise. Betting and pools are for adults 18+, and responsible gambling explains how to keep it under control.