From form to expected goals
Sharpening the estimate with xG
4 min
Raw goals are honest but noisy — a deflected winner or a missed sitter can flatter or flatter-to-deceive a team's record. When it can, the model sharpens its estimate with expected goals (xG).
What xG adds
xG scores every chance by how likely it was to be converted, so it measures the quality of the chances a team created and allowed — not just whether the ball happened to go in. A side out-creating opponents but losing on the scoreboard is often better than its results, and xG sees that where raw goals don't.
How the blend works
When both teams have xG history available, the model computes a parallel expected-goals estimate from xG (the same attack-meets-defence symmetry, but using xG-for and xG-against), then pulls the goals-based figure partway toward it. The xG estimate gets a meaningful but minority weight — it sharpens the number rather than overriding it.
Crucially, this only happens when the data exists. Most league history from the free CSVs carries goals and corners but not xG, so for those matches the model behaves exactly as the previous lesson described — no xG, no change. The blend kicks in mainly where per-fixture statistics have been collected (notably for World Cup teams, covered later).
Why minority weight
xG is sharper but it's still a single-season-ish signal with its own noise, and coverage is uneven. Letting it nudge rather than dominate keeps the model robust: when xG is good it helps, and when it's thin or missing nothing breaks.
xG is a quality lens, not a replacement. It sharpens the goal estimate where it exists and quietly steps aside where it doesn't.