Expected goals (xG) is the single most useful number in football analytics. It answers a simple question: given where and how a shot was taken, how likely was it to become a goal? Here’s how to build a first version from public event data.

The Idea

A goal is a rare, noisy event. A team can play well and lose; a single deflected shot can decide a match. xG smooths out that noise by scoring each chance on its quality rather than its outcome. Sum a team’s xG over a match and you get a better estimate of how many goals they should have scored than the scoreline alone.

The Features That Matter

Most of the predictive power in a basic model comes from a handful of features:

• Distance to goal — the strongest single predictor.
• Angle to goal — how much of the goal mouth the shooter can actually see.
• Body part — header vs. foot.
• Play pattern — open play, set piece, fast break, penalty.

You can compute distance and angle directly from the shot’s (x, y) coordinates in event data. StatsBomb’s open data is a good place to start.

A Minimal Model

import pandas as pd
from sklearn.linear_model import LogisticRegression

# shots: DataFrame with x, y, is_header, is_open_play, goal (0/1)
shots["distance"] = ((120 - shots.x) ** 2 + (40 - shots.y) ** 2) ** 0.5

X = shots[["distance", "is_header", "is_open_play"]]
y = shots["goal"]

model = LogisticRegression(max_iter=1000).fit(X, y)
shots["xg"] = model.predict_proba(X)[:, 1]

Logistic regression is the right starting point: the output is already a probability, the coefficients are interpretable, and it’s hard to overfit. Once this baseline works, gradient-boosted trees usually add a few points of calibration — but get the baseline honest first.

Reading the Output

The number to trust is not any single shot’s xG — it’s the aggregate. Over a season, a striker who consistently outperforms their xG is either an elite finisher or due for regression. Telling those two apart is where the analysis starts, and it’s a theme I’ll keep coming back to.

This is a starter post for the Data & Models track — more on calibration, tracking data, and possession value to come.