Watch enough football, and it’s hard not to develop strong opinions about who can and can’t finish. Some players just seem to hit the ball cleaner, pick the right corner more often and stay calm when a chance falls to them. But watch a striker go through a cold month, and that reputation vanishes overnight — replaced by the sudden certainty that he was never actually good in front of goal to begin with.
I don’t think either of those takes is crazy. Most of us have watched enough football to believe
finishing is a real thing. But if you asked 10 people in the comments what they mean by “finishing,” I’m pretty sure you’d get something like 10 different answers. And before I spend the next few months trying to build a model that measures it, I think it’s worth slowing down and asking what “it” actually is.
Start with a single shot. Say a striker receives the ball just inside the box, slightly off center, with a defender a step behind him. A typical xG model might put that chance at something like 0.10 — meaning a shot from that spot goes in about 1 time in 10. (If you read Fixing xG (Sort Of), you know I have some opinions about how much to trust those numbers, but the idea holds.) Now say he scores. Did he finish well? Probably, in the sense that 90% of those attempts end up as misses. But did he score because he struck it perfectly, because the keeper was slow to react, because it took a tiny nick off a defender’s boot or because it was simply his 1-in-10 moment? From the outside, all four of those look identical on the scoresheet.
That’s the core problem, and it’s worth separating out the pieces. I think a goal is really the product of at least four things stacked on top of each other: the quality of the chance (where the shot comes from, the angle, the pressure, the body part), what the shooter does with it (where he puts the ball, how hard he hits it), what the keeper does in response and plain old luck. xG tries to capture the first one. The rest gets lumped together into the gap between goals and expected goals — which is where most people, myself included, have gone looking for finishing skill.
The obvious definition, then, is that a good finisher is someone who scores more goals than xG says he should. It’s clean and it’s intuitive. It’s also, if you squint at it, a little circular. xG is built by looking at a huge pile of shots and asking how often shots like this one went in. So “expected” doesn’t mean some platonic, neutral baseline — it means “what the players in the training data did.” And the players in the training data include the good finishers. In fact, they probably include a lot of them, since the best finishers tend to take more shots (managers aren’t dumb — if a guy puts the ball in the net, you get him the ball). So we end up defining finishing skill as beating the average of a group that already has the skill we’re trying to measure baked into it. It’s not quite a tautology, but it’s close enough to one that it should make us a little uneasy.
This isn’t just me being pedantic. Jesse Davis and Pieter Robberechts wrote a paper in 2024 arguing pretty much exactly this — that xG models carry biases from how the training data is structured, so the “average player” the model represents isn’t really average, and that this can meaningfully distort finishing skill estimates. It doesn’t break the idea of measuring finishing, but it does mean the ruler we’re measuring with is a little bent.
Then there’s the question of where finishing starts and stops. Mike Goodman wrote a piece for StatsBomb back in 2014 that I keep coming back to, and I think he framed this better than anyone. His argument was basically that people talk past each other because there are two different things we mean by finishing. On one end there’s the narrow version — the pure moment of the foot hitting the ball, with everything else held constant. On the other end there’s the broad version — the ability to beat xG in general, which folds in shot selection, decision-making, weak foot usage and a bunch of other stuff. A player who refuses to shoot from 30 yards and waits for a better chance will look like a better finisher under the broad definition, even if his actual striking is completely ordinary. Is that finishing? I think you could make a case either way, and that’s sort of the point — if we don’t agree on the definition, we can argue forever and never actually disagree about anything.
Even if we agree on a definition, though, there’s a second problem, and this one is about math rather than philosophy. Let’s say a striker takes 80 non-penalty shots in a season, with an average xG of 0.12 per shot. That’s 9.6 expected goals. If he’s a perfectly average finisher — no skill, no curse, nothing — how many goals would he actually score? The honest answer is that anywhere from about 4 to 15 would be totally unremarkable. A normal season for a completely average finisher could look like a disaster or a breakout, and that’s before any actual skill enters the picture.
Now say he genuinely is a good finisher, and he converts 10% more than average. That’s about 1 extra goal a season. One goal of real signal, buried under something like three goals of standard deviation. You’d need several seasons’ worth of shots before that signal really starts to peek out from the noise — and by then the player is older, maybe at a different club, maybe playing a different role. Goodman’s point was that location drives so much of whether a shot goes in that everything else is tiny by comparison, and that makes it really hard to see. I think the numbers back him up.
So that’s where the field has been stuck for about a decade: most people believe finishing skill exists, but it’s small and noisy enough that proving it for any individual player is really hard. The good news is that a lot of smart people have taken swings at it, and the story of those attempts is basically the story of the field getting more careful about the same question.
Around the same time as Goodman, Michael Caley made the case that finishing skill is real — it just only shows up over very large samples. Pool enough shots and the players you’d expect to rise to the top actually do. That’s reassuring, but it also quietly sets up the Davis and Robberechts problem from earlier: if the best finishers are the ones taking the most shots, they’re also the ones shaping the baseline.
Devin Pleuler, writing for Opta, pushed things in a different direction. He split a shot into two stages — the chance itself, and then what the shooter does with the ball once he strikes it — which is really the seed of what we’d now call post-shot xG. (He also later found that even this second-stage measure didn’t carry over very well from one season to the next, which is a caution I’m taking seriously.) More recently, people like Baron and colleagues have looked at where missed shots end up as a source of information, which I think is a clever way of squeezing more out of the data — a shot that just misses the top corner probably tells you something different than one that ends up in the stands.
The other big thread is Bayesian modeling, which sounds intimidating but has a pretty intuitive core. If a player has taken 20 shots and scored four more goals than expected, the model basically says “that’s interesting, but you haven’t earned it yet” and pulls his estimate heavily back toward average. If a player has taken 400 shots and is still +4, the model barely touches him. It’s a formal way of saying “the more evidence you have, the more I believe you.” Sam Gregory, Marek Kwiatkowski at StatsBomb and Laurie Shaw all built versions of this, and more recently Martin Eastwood and Scott Willis have put out finishing models that, I think, are some of the best public work out there. Kwiatkowski in particular flagged shot placement as the obvious next thing to add — which is more or less where I come in.
If I had to boil the whole body of work down, I’d say it’s something like this: finishing skill exists, it’s small, raw goals minus xG is mostly noise over any sample a fan would care about and the tools we use to measure it have their own biases baked in. None of that means it’s not worth measuring. It just means we have to be honest about what we can and can’t see.
So this is the first part of a series on finishing skill — think of it as setting the table before we eat. Over the next few pieces, I’m going to try to build my own finishing skill model, and I’ll be making a few choices that are a little different from what’s been done before. I’ll get into those next time.
I’ll also be honest up front: it’s entirely possible this doesn’t work. My approach might turn out to be just as noisy as goals minus xG or even just conversion rate, in which case the finding is “we still can’t see finishing very well,” and I’ll write that piece too. But I think it’s worth a shot. (Sorry.)
Before part two, I’m curious how deep people want to go on the modeling side. If a more detailed explainer on how the Bayesian piece actually works would be useful, let me know in the comments — I’d much rather over-explain than lose people halfway through.
