Using xG in Value Betting: Build and Test the Feature

Start with a provider definition

Opta defines xG as a shot-level probability estimate built from historical shots and contextual features in its current explainer. Other providers can use different features, labels and model versions, so values are not automatically interchangeable. Open peer-reviewed research also shows that xG model performance depends on feature and modelling choices (Mead, O'Hare and McMenemy, 2023).

Build only lagged features

For a forecast at time t, possible inputs include rolling xG for and against, opponent-adjusted xG, home-away splits and lineup-weighted xG contributions. Every component must use matches and provider records available before t. Corrected event data or final lineups cannot enter an earlier forecast.

Test information gain

Use ordered validation to keep later fixtures outside model fitting:

1. Fit a simple rating or goals baseline.
2. Add one declared xG feature family.
3. Train and tune on earlier periods.
4. Lock the pipeline.
5. Score both models on the same later fixtures.
6. Compare proper scores, calibration, coverage and failure slices.

TimeSeriesSplit documents an ordered validation approach, and calibration guidance supports testing the probabilities rather than only winner accuracy.

Illustrative feature arithmetic

Suppose a team's last five provider xG values before the cutoff are 1.2, 1.8, 0.9, 1.5 and 1.1.

Five-match mean xG = (1.2 + 1.8 + 0.9 + 1.5 + 1.1) / 5 = 1.30. This uses the shot-probability metric described in Opta's xG definition.

That is a descriptive input. It is not the team's expected goals for the next match until a validated model combines opponent, venue, recency and other declared information; peer-reviewed xG research shows that feature and model choices affect the estimate.

Add the market only after forecast validation

Once a model produces a probability for the exact betting outcome, compare it with a timestamped executable price. Keep market probabilities as a baseline and decision input, not a future feature accidentally added to an earlier model. Historical football forecasting research shows why model and market comparisons need out-of-sample evaluation (Goddard, 2004).

Failure conditions

• xG improves only the training period.
• Results disappear under another provider definition.
• Missing lower-league xG is silently filled with zero.
• The feature works only after choosing a window post hoc.
• Forecast improvement does not survive price and execution costs.
• A few dramatic under- or over-performance stories replace complete testing.

Publish the ablation result

Show the baseline and baseline-plus-xG models on the same untouched fixtures. Report log loss, Brier score, calibration, coverage, and uncertainty, then remove each xG feature family in turn. Ordered validation keeps the comparison on later fixtures, while calibration guidance checks reliability rather than only winner classification.

Repeat the comparison after provider updates and on leagues with different event coverage. If the xG model improves aggregate scores but fails when lineups are missing or produces overconfident extremes, do not reduce the result to one “xG works” verdict. Preserve the simpler model as an operational fallback and state the exact conditions under which the feature is released.

Continue learning

• Next guide: Is Value Betting Profitable in 2026? An Evidence Test
• Related guide: World Cup 2026 Dark Horses

Assumptions and limitations

The five-match example is illustrative. xG measures one provider's estimate of shot quality, omits unattempted chances, and is not causal. This page owns the value-model experiment; xG explained covers interpretation in more depth.