Bravsen Predictions: xG-Based Outcome Probabilities

The Crime Index rules on matches after the final whistle. This page runs the same model the other way round: fed with both teams' xG history, it works out the odds of a win, draw or loss, the most probable scoreline, and a metric you won't find on any tipster site — the chance the match itself turns into a crime, a result that contradicts the run of play. Every call is locked in before kickoff and stays on the public record below, misses included.

The sample is still small: across 74 settled calls the confidence interval on that figure runs from roughly 28% to 59%, so it can't yet tell you much about the model. It starts to mean something closer to 150–200 matches. Of those, 1 of 3 come from model v2.

Another 4 calls are already computed for matches further out; they surface here as kickoff approaches. A prediction is frozen when it's made and never recomputed, so it's worth reading close to the match.

Before shipping the model, we ran it backward through the entire World Cup knockout stage under one hard rule: for every fixture, it could only see the xG history that actually existed before kickoff — no hindsight, no peeking at the final score. The result: 24 correct outcomes out of 31, a 77% hit rate. Narrow it to the confident calls, where the favorite sat at 55% or higher, and it climbs to 12 of 14 — 86%.

It got the final right, too: Spain over Argentina at 42% to 29%, built on an expected xG split of 0.87–0.50 — a cautious pick, but the correct one, backed up by the actual 1-0 scoreline and a 20-2 shot count. The two knockout misses tell their own story: Switzerland's scoreless draw with Colombia (combined xG of 0.31 — a coin flip by definition), and the France–England semi-final, a 4-6 thriller that no model on earth would have called. Thirty-one matches is a small sample, and we're upfront about that: the real exam started with the club season, and its results are accumulating in the track record above.

That exam has proved harder than the backtest, as it should. The World Cup knockouts featured teams with long match histories and wide gaps in class; a club season starts from scratch, with four to six matches of data per side and squads reshaped over the summer. The first few dozen club calls landed near 42% against 77% in the tournament backtest, and both numbers stay on this page for exactly that reason: the backtest shows the model's ceiling on convenient data, the track record shows what it does on inconvenient data.

No black box here. A team's attacking strength is the average xG it has produced across its last ten matches with real tracking data; its defensive weakness is the average xG it has conceded over the same run. Expected goals for the upcoming fixture come from multiplying one side's attack by the other's leakiness, adjusted for the league's scoring level and a home-field bump (dropped entirely on neutral-venue tournaments). A Poisson distribution then unpacks those two numbers into a full matrix of scoreline probabilities, from which we read off win/draw/loss odds and the single most likely score.

Anomaly risk is the probability of a result going against the xG favorite, boosted for "small-numbers football": the lower the combined expected goals, the more a match hinges on one bounce — the core lesson from our xG explainer. Each prediction is calculated once and frozen — recalculating the odds an hour before kickoff and then bragging about accuracy isn't something we're willing to do. If a team has fewer than four matches of xG data, the model stays silent. Bravsen doesn't publish odds and doesn't give betting advice — this is statistics, not bookmaking.