How LMEX Makes Football Predictions

Overview: what a prediction is

A football prediction from LMEX is the output of a fixed rule that works on numbers. It is not one person's opinion. Give the rule the same data and it gives the same answer every time. This page is written so that you can check how we reach a prediction and see the limits of the method. Every prediction is a "Statistics-based opinion, not a guarantee · 18+". Football is highly uncertain and we never say that any result is certain.

The aim of the rule is not to guess who wins. It looks for the side that is easier to back than the market price says. If the market gives the home side a 50% chance to cover its handicap and our statistics give a clearly different figure, we pick the home side. If there is no big enough difference we write "no prediction", which is just as common an outcome as a prediction.

The data we use

About a day before kick-off the system fetches analysis data for the match from a football data provider, and again in the last three hours before kick-off. Only matches played before the match being analysed are counted.

We do not use team news, line-ups, injuries or gut feeling, because that information is not in our data set.

The steps of the rule

Step 1: expected goals for each team. The home value is the average of the goals the home team scores per match and the goals the away team concedes per match. The away value is worked out the same way the other way round. When there are three or more head-to-head meetings, use a 75% weight on the form value and 25% on that team's average goals in the head-to-head meetings. Then multiply by 1.06 for the home side and 0.94 for the away side to allow for playing at home, and keep every value between 0.35 and 3.5 goals. If either team has fewer than three matches of data, the match is not assessed at all.

Step 2: the probability of every score. We assume each team's goals follow a Poisson distribution with the averages from step 1 and that the two teams score independently. That gives a probability for every score from 0-0 to 10-10.

Step 3: the chance to cover the market line. For the latest handicap line and the latest over/under line we add up the probabilities of the scores that make each side win. Quarter lines (.25 and .75) are treated as the half win or half loss they really are: a half win counts 0.75 and a half loss counts 0.25. A push (stake returned) is left out of the calculation.

Step 4: remove the margin from the market price. The two prices of a market are converted to decimal odds (add 1 to the Hong Kong price). The probability of each side is 1 divided by its odds, and the two are scaled to add up to 100%. That is the probability the market really implies, which we call the fair price.

Step 5: pull towards the market. The market knows far more than ten matches of form, so we pull the step 3 value halfway towards the market: final value = fair price + 0.5 x (model value - fair price). The gap between the final value and the fair price is called the edge.

Step 6: pick or pass. Of the four possible sides (home or away on the handicap, over or under on the goals line) we take the one with the biggest edge. If the edge is at least 6 percentage points (the raw model value is at least 12 points away from the market) we publish a prediction. If the raw model value is more than 25 percentage points away from the market we publish nothing, because that usually means the data of the two teams cannot be compared, for example a cup tie between clubs from different divisions. If the edge is smaller than the threshold we write "no prediction" together with the reason.

A worked example

Suppose the home side averages 1.9 goals scored and 0.9 conceded and the away side averages 1.0 scored and 1.6 conceded. The home value is (1.9 + 1.6) / 2 = 1.75, times 1.06 gives 1.86 goals. The away value is (1.0 + 0.9) / 2 = 0.95, times 0.94 gives 0.89 goals. The market has the home side giving 0.25 of a goal at prices of 0.90 and 0.94. After removing the margin the market gives the home side 50.5% to cover, while the model calculates 65.6%. After pulling halfway towards the market that falls to 58.1%, an edge of 7.5 points. This is above the 6-point threshold and the raw gap of 15 points is below the 25-point limit, so the system publishes "home -0.25 at 0.90". All numbers here are an illustration, not a real match.

What happens after publication

A published prediction is frozen: the line, the price and the time cannot be edited or deleted. If the first assessment, about a day before kick-off, gave "no prediction", the system assesses the match again once the last data before kick-off arrives and publishes a prediction if it then passes the threshold. After kick-off nothing changes.

When the match ends, every prediction is settled on the full-time score. There are five possible results: win, half win, push, half loss and loss. Predictions for matches that are cancelled or postponed by more than three days are counted as void and left out of the record. Profit and loss assume a stake of 1 unit on every prediction at the published price: a win pays the price, a half win pays half the price, a half loss loses half a unit and a loss loses one unit. The full record is on the football predictions today page.

Limits you should know about

The rule is deliberately simple. It cannot see team news, does not know about injuries or suspensions, does not separate home form from away form and weights the last five and the last ten matches equally. An Asian handicap market has a hidden margin of roughly 3% to 5%, which means a prediction can win more than half the time and still lose money, and short runs of results swing widely. We have not tested the rule on a large amount of historical data. The public record on the predictions page is the only evidence we have and we will keep adding to it.

If the record shows that the rule does not work, we will say so plainly and change or retire the rule, without deleting the old predictions. Gambling is for adults aged 18 and over. Read the full policy and where to find help in the editorial and responsible gambling policy.