Statarea predictions, and the cost of weighting recent form

What recency weighting actually does

Every form-based model has to answer one question: how far back does form count? A model that averages the whole season treats a match from August the same as a match from last week. A model that only looks at the last three matches has almost no data. Recency weighting is the compromise, and Statarea sits deliberately towards the responsive end.

The case for it is real. Football teams are not fixed objects. A new manager arrives, a first-choice striker comes back from injury, a defensive problem gets fixed. When something like that happens, the season average is describing a team that no longer exists, and a recency-weighted model notices weeks earlier.

The case against it is equally real and less often stated. Consider a mid-table side that has won three in a row. How much of that is improvement and how much is variance? In a sport where a single deflection decides matches, three results is nowhere near enough to separate the two. A model that leans on those three matches will treat a lucky run as a step change, and it will do so confidently, because the arithmetic does not know the difference between signal and noise. It only knows what it was told to weight.

The sample-size problem, with numbers

This is worth making concrete, because “small sample” is easy to nod along to and hard to feel.

Take a team whose true chance of winning any given match is 40%. Over three matches, the chance it wins all three by luck alone is 0.4 x 0.4 x 0.4, which is 6.4%. That is not a freak event. Across a division of twenty clubs playing simultaneously, several teams will be on a three-match run at any moment purely by chance, and none of them will have improved at all.

Now flip it. That same 40% team loses three in a row with probability 0.6 x 0.6 x 0.6, which is 21.6%. More than one time in five. A model weighting the last three heavily will mark that team down sharply, and four times out of five nothing has actually changed.

This is the same arithmetic that makes a short run of correct picks meaningless as evidence about a tipster, which is why our own record asks you to wait for a large sample before giving it any weight.

When to lean on it and when to discount it

• Lean on it when form and structure agree. If recent results, the league table and the head-to-head all point the same way, the weighting is amplifying signal rather than noise.
• Discount it after a manager change. Not because the model is wrong to react, but because the data on which it is reacting is the two or three matches since the change, which is the smallest sample it will ever work from.
• Discount it heavily in cup ties. League form predicts league selection. A cup tie is often a different eleven with different motivation, and no amount of weighting fixes an input that does not apply.
• Discount it in thin leagues. Recency weighting needs a reliable baseline to be recent relative to. Where the underlying data is sparse, the weighting simply makes a noisy estimate noisier.
• Watch for the fixture list behind the form. Three wins against the bottom three is not the same as three wins against the top three, and a pure form sequence does not distinguish them.

Credit where it is due: it shows the inputs

Statarea puts the form data on the match page rather than only the conclusion. That matters more than it might seem. A site that gives you a bare selection is asking for trust; a site that gives you the numbers it used is inviting you to disagree with it.

Use that. If the prediction rests on a run of results against weak opposition, or on a head-to-head record from squads that have since turned over completely, you can see that on the page and discount accordingly. The information to overrule the model is right there, which is not true of most free tips.

What is missing is one level up: not the inputs to a single forecast, but the outcomes across all of them. There is no published season-long settled record with prices attached, so the question of how the weighting has actually performed cannot be answered by a reader. That gap is standard across the category, and Forebet and PredictZ have it too.

Using it on a Kenyan slip

Most Kenyan readers arrive at a stats site on the way to a jackpot, so the practical warning is worth stating directly. Jackpot slates are built from whatever is playing, which in practice means a lot of second divisions and a lot of leagues you have never watched. Those are the conditions in which a recency-weighted form model is at its least reliable.

The sensible use is to treat the confident-looking predictions in major leagues as worth weighing, and to treat the equally confident-looking predictions in obscure competitions as barely better than the base rates. The model presents both in the same format with the same authority. The difference in how much they are worth is entirely invisible on the page, and it is large. The arithmetic of what that does to a 17-leg slip is on how jackpots actually work.