Football score predictors: what they can and cannot do

Search for a football score predictor and you get three very different things wearing the same label. Telling them apart is most of the skill, because one of them is genuinely useful and the other two are confident noise.

The three kinds

A rate model. Estimates how many goals each side should score, then converts that into a probability for every scoreline. It will happily tell you the most likely result is 1:1 at 12 percent, which sounds unhelpful and is actually the honest answer. The formula behind it is worth ten minutes of your time.

A form widget. Takes the last five results, applies a rule of thumb, prints a score. No probabilities, no calibration, no way of knowing whether it has ever been right. These are the ones that come bundled with fixture listings.

A chatbot. Ask a language model for Saturday's scores and it will give you ten of them, in a neat table, with reasons. It is doing something closer to writing than forecasting, and it has no idea who is injured this week unless it went and looked.

How to tell whether one is any good

• Does it show probabilities, not just a score? A predictor that says 2:1 with no number attached is hiding the fact that 2:1 is a one in ten shot.
• Is its record public and dated? Predictions published after kickoff are worthless. So are records you cannot check.
• Does it say how often it expects to be wrong? The good ones lead with it. Nobody beats about 15 percent on exact scores.
• Does it change when the team news does? If the number is identical on Monday and on Saturday morning, it is not using the information that moves matches.

A calibrated model that is right 12 percent of the time is better than a confident one that is right 12 percent of the time and never admits it.

The thing every predictor misses

Models run on what has already happened. The matches where they help least are the ones where the past is least relevant: a derby, a manager's first game, a side already safe playing one chasing a European place, a squad that played in another country on Thursday.

Those are exactly the matches worth thinking hardest about, and they are why our multiplier goes up when the numbers get murky rather than down. How the multiplier is calculated is public, and so is the reasoning behind the hardest fixtures on a coupon.

Using a predictor in a prediction league

Here is the trap. If a tool is any good, it is giving the same answer to everyone who asks it, and its answer is the crowded pick. Copy it across ten matches and your coupon is the average coupon in your league. You will finish mid-table forever, being right about as often as everyone else and gaining ground on nobody.

The way to use one is as a baseline you argue with. Take the ranked list, find the two or three matches where you have a real reason to disagree, and back yourself there. That is where positions in a league actually change hands, and it is why correct scores nobody else called are worth a rarity bonus in our scoring.

If you already keep your picks somewhere else

Plenty of people work their round out in a notes app, a group chat, or a conversation with a chatbot, then have to type it all in again. We built the round page so you do not have to: there is a paste box that reads a block of text, matches the team names to the fixtures, and fills the scores in for you. Paste a list, paste a table, paste a chatbot's answer with all its waffle attached, and it picks out the scorelines. Nothing is saved until you press save, so you can fix whatever it misread first.

That is the useful division of labour. Let the tool do the arithmetic and the typing. Keep the disagreeing for yourself, because the disagreeing is the game.