I've been writing previews and predictions for a long time and I want to start with an admission that isn't good for business: I get a lot of them wrong. So does everybody else. So do the models, so do the bookmakers, and so does the person who was extremely confident on the podcast last week.

This isn't false modesty. It's a structural feature of the sport, and once you understand it properly you start making better decisions about what's worth predicting at all.

The low-scoring problem

The single biggest reason football is hard to predict is that goals are rare.

A typical match produces two or three goals. Compare that to basketball, where a hundred-plus points per team means individual events average out and the better team usually wins. In football, a single deflection, a single refereeing decision, a single moment of goalkeeping can decide everything.

The consequence is that the gap between team quality and match outcome is enormous. If you simulate a match between a clearly superior side and a clearly inferior one thousands of times, the better team might win 60% of the time. Not 90%. Sixty.

Which means that in any single match, the "wrong" result happens constantly, and it isn't an upset in any meaningful sense. It's the expected frequency of a coin that lands heads slightly more often.

Draws break everything

The three-outcome structure makes it worse. Roughly a quarter of matches end level, and the draw is the outcome that's hardest to predict specifically because it's not associated with either team being better.

A draw arises from a particular kind of match — evenly matched, or one side defending well, or both sides being cautious. Those conditions are real but they're not strongly predictable from team quality, which is what most prediction inputs measure.

This is why almost nobody predicts draws, and why prediction accuracy figures look worse than they should. Predicting a home win in every match would produce a respectable-looking hit rate and would be completely uninformative.

What actually carries signal

Having established that certainty is unavailable, some things genuinely do predict better than others.

Underlying performance over a long window. Shot quality created and conceded over fifteen or twenty matches predicts future results better than actual results over the same period. Teams that create good chances and concede few tend to keep doing so; teams winning without doing either tend to stop.

Squad availability, but only for key players. Losing one important player has a measurable effect. Losing a squad player has almost none, and team news pages treat both identically.

Style matchups. A team that struggles against a high press will struggle against every team that presses high. These effects are real, reasonably persistent, and much less discussed than form.

Rest and travel. A side playing its third match in seven days, having travelled, against one that's had a week off. This is one of the more reliably predictive factors and it's frequently ignored in favour of narrative.

What carries almost none

Recent form. Five or six matches is far too small a sample. A team that's won four in a row has usually been slightly lucky, and regression is more likely than continuation.

Head-to-head history. Different players, different managers, different circumstances. There's an exception for genuine derbies and I'll come to that elsewhere, but as a general rule it's noise.

Motivation narratives. "They need this more." "They've got nothing to play for." These make good copy and they're extremely weak as predictors. Professional footballers mostly try.

Manager pressure stories. The new-manager bounce is real but small, and considerably smaller than the folklore suggests.

The right way to hold a prediction

The framing that helped me most was moving from picking outcomes to estimating probabilities.

Instead of "Team A will win," think "Team A wins about 45% of the time, draw 28%, Team B 27%." That's a much more honest statement and it's actually testable — if you make a hundred predictions at 45% and they come in 43% of the time, you're calibrated, which is the only meaningful measure of prediction skill.

It also changes how you evaluate yourself. A prediction that didn't come off wasn't necessarily wrong. If you said 45% and the other thing happened, that's the 55% arriving, which it does more often than not.

The corollary is that judging predictions by hit rate is nearly meaningless. Somebody who only predicts heavy favourites will look brilliant and will have said nothing.

Where prediction is actually possible

Some things are far more predictable than match results and get much less attention.

Season-long league positions, over 38 matches, are quite predictable, because the sample is large enough for quality to dominate variance. Most preseason predictions of the top four are broadly right.

Total goals in a match is more predictable than the winner, because it depends on both teams' styles rather than on which of two similar sides gets a decisive moment.

And relegation and promotion battles resolve towards underlying quality with remarkable consistency once you're twenty-odd games in.

The honest position

If somebody tells you they know what's going to happen in a specific football match, they're either selling something or they haven't thought about it much.

What a good preview can do is tell you what to watch for — which matchups matter, where the game is likely to be decided, what each side is trying to do. That's genuinely useful and it doesn't require pretending to know the score.

The prediction at the bottom is mostly there because readers expect one. I'd suggest treating it as the least interesting part of anything you read, including this.