Expected goals (xG) explained for betting

In one sentence

Expected goals (xG) is the probability that a shot becomes a goal, estimated from thousands of similar past shots, and summed to measure the quality of chances a team creates.

How it works

Goals are rare and luck plays a big part: a team can dominate and lose 1-0. xG looks past the scoreline to the chances. A penalty is worth about three-quarters of a goal; a speculative 30-yard shot might be worth 0.03.

Each shot's value comes from a model, usually logistic regression or gradient boosting, trained on past shots. Inputs include distance, angle, body part, type of assist and whether it was a set piece. Richer versions add defender positions and goalkeeper location.

For betting, xG totals over several matches are a steadier guide to attacking and defensive strength than goals, and they feed goal models that price over/under and match odds.

The maths

xG_{\text{shot}} = \frac{1}{1 + e^{-(\beta_0 + \beta_1 x_1 + \dots + \beta_k x_k)}}

xG_{\text{team}} = \sum_{\text{shots}} xG_{\text{shot}}, \qquad P(\text{at least one goal}) = 1 - \prod_{\text{shots}} (1 - xG_{\text{shot}})

x_1 to x_k: shot features such as distance, angle and header or foot.
β values: weights learned from historical shots.
xG_team: expected goals from all the team's shots.
The product term: the chance every shot misses, assuming shots are independent.

In words: each shot gets a scoring probability from its features, and adding them up tells you how many goals the chances were worth.

Worked betting example

Part one, a single match. A team takes five shots worth 0.76 (penalty), 0.12, 0.05, 0.31 and 0.08.

Team xG = 0.76 + 0.12 + 0.05 + 0.31 + 0.08 = 1.32.
Chance of scoring at least once = 1 − (0.24 × 0.88 × 0.95 × 0.69 × 0.92) ≈ 87.3%.

Part two, pricing. From recent xG form, you estimate the home side will generate 1.55 expected goals and the away side 1.15.

Total expected goals λ = 2.70. Treating goals as Poisson, P(3 or more goals) ≈ 50.6%. Fair price for over 2.5 ≈ 1.97.
Betfair offers 2.10 on over 2.5. £10 stake, 2% commission: a win pays £11 × 0.98 = £10.78. EV ≈ 0.506 × £10.78 − 0.494 × £10 ≈ +£0.51.

A small edge like this can vanish with a modest error in either team's xG estimate.

Where it's good

Rating teams early in a season when goal tallies are dominated by luck.
Spotting teams whose results outrun their chances, a common sign of regression to come.
Feeding Poisson and Dixon-Coles models with steadier attack and defence inputs.
In-play, judging whether a scoreline reflects the run of play.
Player scorer markets, where shot volume and quality matter more than recent goals.

Limitations and pitfalls

Different providers give different xG for the same shot; do not mix sources.
Basic models ignore defender and keeper positions, so they undervalue open goals and overvalue crowded shots.
xG ignores chances with no shot, such as a cross that just evades a striker.
Finishing skill exists for some elite players, but it is small and hard to prove from limited shots.
Game state distorts xG: a team leading 2-0 may sit back and concede many low-value shots.
xG is now mainstream and bookmakers and exchange traders use it, so obvious xG-versus-goals gaps are often already priced.
Summing independent shot probabilities misstates rebounds; many models group shots from the same move.

How to build it

Python: scikit-learn or lightgbm for the shot model, statsmodels for logistic regression; mplsoccer for plotting.
Data: event data with shot locations and types, from providers such as StatsBomb (some free open data) or Opta.
Check calibration: across all shots rated around 0.10, about 10% should be goals.
Tip: use rolling, time-weighted xG for and against, and blend with actual goals rather than discarding them.

Logistic regression: the classic way to build a shot model.
Poisson distribution: turns xG into scoreline probabilities.
Dixon-Coles model: a full scoreline model that can use xG inputs.
Gradient boosting: a common modern shot model.
Pi-ratings: can be driven by xG difference instead of goals.