My xG vs the supercomputer: what Opta can't see

Opta's xG 2.0 takes into account more than 35 contextual factors — shooter position, goalkeeper position, defensive pressure, shot clarity. It's a genuine supercomputer and I'm not arguing with it: a model trained on thousands of shots is more consistent than any human, it doesn't get tired and it doesn't get emotional.

But a model has one fundamental limit: it rates shots. Football is much bigger than shots. I've been building my own xG model for years and I've concluded that four things make the biggest difference: penalties, chances without a shot, razor-thin offsides, and game state.

Opta gives every penalty a flat value of 0.79 (StatsBomb 0.78, Wyscout 0.76) — the historical conversion rate, since the shot itself is always the same: a stationary ball, 12 yards, little variation. As a shot, that's right.

But for me xG exists to measure a team's performance, not a single shot. Two penalties can come from completely different places. An accidental handball off a cross into the box is almost a matter of luck. A last-man foul that stops a certain goal, on the other hand, is the result of the team creating a real chance. The first doesn't reflect how the team played; the second does.

So I don't automatically give a penalty the same value — I judge the situation: how did it come about? That way what enters my xG is what the team actually created, not what the referee awarded.

By definition xG is tied to a shot: no shot, no xG. Now think of a perfectly weighted through-ball across the six-yard box where the striker is one stride from touching it. Or a cross that beats everyone and reaches the striker at the far post, who can't get there. These are often exactly the situations that come out of a well-working move, and they're genuine big chances — yet in the xG table their value is zero.

A team that creates such situations systematically looks worse in the data than it really is. This is where a human who watches the game is ahead of the model.

Take a situation where a team breaks the defensive line with a perfectly timed run, but the flag goes up for a very tight offside. In the model there's nothing: no shot was taken or the goal doesn't count, so xG is zero. Yet the team did exactly what it's supposed to do — got behind the defence, played the right pass, timed the run. No table takes that into account.

In betting, that can be a valuable signal. One such moment is chance — a centimetre on the line. But if there are several in one match, it shows the team consistently finds the space behind the defence, and that's a good sign of performance the numbers don't show. Matches where xG underrates a team are exactly where the market can be wrong.

The same xG isn't worth the same at every moment of a match. Three examples of how I'd adjust it:

A trailing team's xG could carry a smaller multiplier (say 0.75×). They have to equalise, they take risks, shoot more and often from worse positions while the opponent sits deep — the same xG is “cheaper” here.

An underdog's xG could carry a larger multiplier (say 1.25×). xG doesn't measure underdogs' chances fairly: in reality they sometimes get more than half of a match, but the model rates their chances by average-team standards and underrates them.

A leading team's xG could also carry a larger multiplier. They have no specific motivation to score more, so every chance they do create is a stronger sign of their real level. It depends on the match, of course.

The market — live too — reacts to the score and to public xG. If public xG says team A had 2.1 xG vs B's 0.6 but half of it came while trailing in the last 20 minutes, the real gap is smaller than the number suggests. And vice versa: if a leading team created little but every chance was top quality, that's a stronger signal than the number lets on.

Those are exactly the spots I look for — where my read diverges from public xG — because that's where the price can be wrong.

The 0.75× and 1.25× multipliers are examples of my thinking, not laboratory-calibrated constants. To really work they'd need testing: does adjusted xG predict the next match's result better than raw xG? The supercomputer's edge is consistency and scale; mine is context it can't see. The best result comes from combining them: the model gives the base, I add the situation.

And these are just examples. Once you start thinking, you find more and more of these situations — a red card, a lineup after an injury, a match where one team has nothing left to play for. Every such nuance is a place where context beats the raw number.

Opta Analyst, “What Is Expected Goals (xG)?”; Stats Perform, “Expected Goals (xG): The Football Metric Changing Analysis, Betting, and Fan Engagement”; Jobs in Football, “What Is The xG Of A Penalty?”. The multipliers and the treatment of penalties are the author's own methodology, not Opta's or other data providers'.