Forebet predictions, and how to read them

What the model is actually doing

Forebet belongs to a family of models built on the Poisson distribution, which is the standard statistical tool for counting events that happen at some average rate over a fixed period. Goals in a football match fit that shape reasonably well.

The mechanism, stripped of jargon: estimate how many goals the home side tends to score against opposition of this quality, and how many the away side tends to concede. Combine them into an expected goals figure for each team. From those two numbers you can compute the probability of every plausible scoreline, 1-0, 2-1, 0-0 and so on. Add up all the scorelines where the home team wins and you have a home-win probability. Adjust for home advantage and recent form, and that is the forecast.

Two things follow from this that are worth holding onto. First, the model has no idea who is injured unless injury data is fed in, no idea that a cup tie is being used to rest players, and no idea that a squad landed at 3am. Second, it is working from historical rates, so it is strongest where there is a lot of history and weakest where there is little. That is why performance varies so much by league, and why a model that looks respectable on the Premier League can be close to guessing in a competition it has thin data for.

Probabilities are a better output than tips

Give Forebet credit for the format. A site that tells you “home win, 54%” is telling you something a site that says “BANKER: HOME WIN” is deliberately hiding, which is that the other 46% exists.

A 54% forecast that loses is not a wrong forecast. It is a forecast that landed on the other side of a coin it told you was nearly fair. You can only judge probabilistic predictions across hundreds of them, by asking whether the things called 54% actually happen about 54% of the time. That test is called calibration, and it is the correct way to assess any forecaster, including us.

Which brings us to the gap.

The number nobody in this category publishes

To run a calibration test you need the full history of forecasts, settled against results, with nothing removed. No major prediction site aimed at this market publishes that, and Forebet does not either. There are no public Brier scores, no backtests, no season-long settled record, and no documentation of what goes into the model or how it is weighted.

So when you see “Forebet is 62% accurate” or “85% accurate”, ask the same question of both: 62% of what, counted by whom, over which matches, in which markets, across what period, and with which results excluded? The range between those two figures is far too wide for both to be measuring the same thing. At least one of them is counting selectively, and you cannot tell which because the underlying data is not published.

This is exactly why our accuracy record exists and why it shows the losses. It is not a claim that we are better forecasters than a statistical model. It is the only format in which such a claim could ever be checked.

Why hit rate is the wrong number anyway

Suppose a model really is right 62% of the time on 1X2. That sounds strong. It tells you almost nothing about whether following it makes money.

A model reaches a high hit rate most easily by picking heavy favourites, and heavy favourites are priced accordingly. Being right 62% of the time at average odds of 1.40 returns 0.87 for every 1.00 staked, a steady loss. Being right 40% of the time at average odds of 3.00 returns 1.20, a profit. The hit rate went down and the outcome went up.

Any accuracy figure quoted without the odds attached is showing you the flattering half of the picture. That applies to Forebet, to every rival site in the same category, and to us. The maths behind it is on our odds explained page.

How to use it sensibly

• Read the probability, not the pick. The percentage is the information. The bolded selection is that percentage with its uncertainty deleted.
• Compare it against the price. Convert the bookmaker's odds into an implied probability and see whether the model disagrees enough to matter. If the model says 54% and the price implies 53%, there is nothing there.
• Distrust it most where data is thin. Lower divisions and unfamiliar competitions are where a statistical model is weakest, and they are heavily represented in Kenyan jackpot slates.
• Add what the model cannot see. Team news, rotation, fixture congestion and motivation are the inputs a goals-based model structurally lacks.
• Never treat any of it as certainty. Nothing that reaches you through a public website is a sure bet, including a well-constructed probability.

Where we differ

We are not a model. We publish fewer selections, with the reasoning written out, and we settle every one of them publicly whether it won or lost. That is a different product from a computed forecast across hundreds of leagues, and it is worse than a model at scale and better than a model at being checkable.

Use both, if you like. But apply the same test to us that we have applied to Forebet here: what does this site publish that would let me catch it being wrong? For us the answer is the record. If a prediction site cannot answer that question at all, that is the finding.