"They're due a goal" is not a model. A field guide to the expected-goals talking points that fall apart on contact with the data.
Expected goals has gone from niche metric to matchday furniture — quoted on broadcasts, argued about in the group chat, and misunderstood in roughly equal measure. Most of the misunderstanding comes from a handful of talking points that sound clever precisely because they're wrong in a comfortable way. Here are five worth retiring.
Finishing variance is real, but 'due' implies a debt the universe repays. It does not. A striker under-performing his xG is telling you about sample size, not writing a promissory note. Over 400 shots the noise mostly washes out; over 12 it dominates. When a pundit says a forward is due, they've quietly assumed the last ten games predict the next one. They don't.
This one survives because it's almost true and completely useless. Yes, more shots usually means more xG. But a model that only counted shots would rate a hopeful 30-yard drive the same as a tap-in. The entire point of xG is that location, body part, assist type and defensive pressure move a chance from 0.03 to 0.7. Teams that out-shoot and under-xG their opponents are the ones the table is about to punish.
A 0.75 xG chance is not a goal you were robbed of. It's a coin that lands your way three times in four. Miss it and the outrage is real; the expectation was never certainty. Judging a striker on a single spurned big chance is judging a poker player on one river card.
Underdogs who sit deep and counter genuinely do generate fewer, better chances — and a good model rewards exactly that. What xG won't do is flatter a smash-and-grab where a side rode 0.4 xG to three points. That isn't the model hating you; that's the model remembering that Tuesday's miracle is not Saturday's plan.
“Expected goals doesn't tell you who deserved to win. It tells you who is going to keep winning.”
— Dele Adeyemi
A 22% chance comes in roughly one time in five. When it does, the model wasn't wrong — it was calibrated. The only fair test of a prediction is whether its stated confidences match reality across hundreds of calls, which is exactly why we publish our calibration score. Anyone grading a probabilistic model on a single result has already lost the plot.
None of this makes xG the last word. It's a lens, not a verdict — best read alongside the tape, the team news, and yes, the occasional gut feeling. Just don't let the clever-sounding version of the argument win the group chat.