Why building your own xG model gives you an edge when betting
You already know basic statistics—shots, goals, and possession—don’t tell the whole story. Expected goals (xG) measures the quality of chances, not just quantity, and can reveal value where market prices lag. By building your own xG model, you control the inputs, correct biases, and tailor the output to betting scenarios you care about (over/under, match result, player props). That means you can spot mispriced odds and act before the market adjusts.
Rather than treating xG as a black box, you’ll benefit from understanding how it’s constructed: what data matters, how features are engineered, and which modeling choices produce stable probability estimates. In this first part you’ll focus on data sources, essential features, and simple modeling frameworks you can implement and iterate on.
What event-level data you need and where to find it
Your model’s quality starts with data. At minimum, collect shot-level events with these fields:
- Shot timestamp and match context (minute, home/away, scoreline)
- Shot coordinates (x, y) or distance and angle to goal
- Body part and shot type (foot, header, own-goal excluded)
- Assist type / buildup (through ball, cross, set piece)
- Defensive pressure indicators when available (marked/unmarked)
- Outcome (goal/no goal, saved, blocked, off-target)
Commercial providers (Opta, StatsBomb, Wyscout) and public datasets (Understat, FBref exports, StatsBomb open) are common sources. If you have limited budget, start with publicly available event feeds or scraped shot coordinates and then scale up once your pipeline is stable.
Early feature engineering: what to compute first
Raw coordinates aren’t enough. You should derive features that directly affect shot quality. Prioritize these transformations:
- Distance to goal and shooting angle — compute using coordinates; these are the most predictive single features.
- Shot type flags — penalty, free kick, header, volley, rebound, or set-piece.
- Build-up context — number of passes in the sequence, whether the chance came from open play or a set piece.
- Game state features — match minute, score difference, home/away; late-game shots under a trailing team bias can shift probabilities.
- Player and goalkeeper context — shooter’s historical finishing stats and the goalkeeper’s save rate can be included if you want personalized xG.
Also consider spatial binning: group locations into small zones (e.g., hex grids or 1–2 m bins) to capture non-linearities in shot success across the box. Keep an eye on multicollinearity (distance and angle are related) and standardize numerical features before modeling.
With clean, well-engineered features, you can move from descriptive xG to predictive xG. The next step is choosing modeling algorithms, training and validating on out-of-sample data, and calibrating probabilities so they align with real-world scoring rates—topics that we’ll cover in the next section.
Choosing and validating modeling approaches
Start with a simple, interpretable baseline and only increase complexity when it improves out-of-sample performance. A logistic regression (or regularized logistic such as L2/Ridge) on distance, angle, shot type and a handful of context features is a robust starting point: it’s fast, easy to calibrate and gives clear coefficients you can sanity-check. From there, test tree-based models (XGBoost, LightGBM) for non-linear interactions and small neural nets if you have lots of data or want to incorporate embeddings (player IDs, location grids).
Validation is where models fail most often. Avoid random shot-level splits that leak temporal or match context. Prefer:
– Temporal splits: train on earlier seasons, validate on later seasons to mimic live deployment.
– Match-aware folds: ensure shots from the same match (and preferably same team) don’t appear in both train and test folds.
Use log loss and Brier score as primary optimization metrics (they penalize probability errors), and use AUC/ROC only as a secondary check of discrimination. Plot reliability diagrams (calibration curves) and compute sharpness to understand if your predicted probabilities are both accurate and useful.
Address class imbalance (goals are rare) with proper weighting or by optimizing log loss directly. Regularize aggressively for high-cardinality categorical features (player/goalkeeper IDs) and prefer target encoding with cross-fold averaging to avoid leakage.
Calibration matters more for betting than raw ranking. If your model is overconfident, apply post-hoc calibration — Platt scaling (logistic regression on model outputs) or isotonic regression — and re-check on holdout sets. Ensemble approaches (stacking a calibrated meta-model on top of base learners) often give the best trade-off between accuracy and well-calibrated probabilities.
Turning xG outputs into betting signals and testing profitability
Shot-level xG becomes actionable once you aggregate it to match and market-level quantities. Common methods:
– Sum shot xG by team to estimate expected goals for the match. Feed team xG into an independent or bivariate Poisson model to derive scoreline probabilities and implied match odds.
– Simulate the match directly by treating each historical or generated shot event as a Bernoulli with its xG and running Monte Carlo simulations to capture sequencing effects (substitutions, red cards can be simulated too if modeled).
– For player props, adjust shot-level xG by minutes played and substitution tendencies; normalize for starter impact and minutes-per-start.
Compare your model-implied odds to market prices using expected value (EV = model_prob * market_payout – (1 – model_prob)). Backtest on historical markets using archived closing and pre-match odds (Pinnacle, Betfair, odds archives). Important practices:
– Use only odds available before the market would reasonably react to new information (e.g., before team news).
– Account for bookmaker margin and liquidity constraints.
– Track unit ROI, strike rate, and maximum drawdown per market type (moneyline, over/under, player props).
Finally, adopt disciplined staking. Flat stakes are simple and reduce risk of overfitting to variance; the Kelly criterion can maximize growth but requires conservative shrinkage and robust edge estimates to avoid large drawdowns. Continuously monitor model performance and rebalance: a tiny calibration drift can turn positive EV into losses when stakes are scaled.
Deploying and evolving your xG system
Once you have a validated model and a tested betting strategy, focus on reliable delivery and continual improvement. Automate your data pipeline and model training so new matches and lineups feed into predictions with minimal delay. Log predictions, odds taken, and outcomes for every bet to enable rigorous performance analysis and to detect calibration drift early.
Operational considerations matter as much as model accuracy: implement rate limits when scraping odds, handle missing or delayed event feeds gracefully, and maintain version control for models and feature code so you can roll back if a change degrades performance. Use monitoring alerts for unusual shifts in matchup-level edges, sudden drops in ROI, or market behavior that suggests information leakage.
Ethics and risk management are integral. Scale stakes conservatively, expect long variance runs, and diversify across markets (match result, totals, player props) to reduce dependency on any single signal. If you’re sourcing free or community data, cross-check it against higher-quality commercial feeds periodically—open datasets like StatsBomb open data are excellent references for benchmarking and testing.
Frequently Asked Questions
How much historical shot data do I need to build a reliable xG model?
More data is generally better because goals are rare events and models need many examples to estimate probabilities reliably. A reasonable starting point is several thousand shots (covering multiple seasons). If you plan to include player-specific effects, you’ll need much larger samples or to regularize heavily to avoid noisy estimates.
Can a custom xG model consistently beat betting markets?
Yes, but not guaranteed. A bespoke xG model can uncover edges by correcting biases in public models or exploiting niche markets (player props, lower-league games). Profitability depends on data quality, realistic accounting for bookmaker margins and market timing, disciplined staking, and the ability to act on edges before the market adjusts.
How do I prevent overfitting when adding player and goalkeeper features?
Use strong regularization, target encoding with out-of-fold averages, or embeddings with dropout. Prefer hierarchical or blended estimates that combine player-level metrics with league or position-level priors. Validate using temporal or match-aware splits to ensure improvements generalize to future matches rather than reflecting noise.
