Sports Analytics

Sport generates a stream of structured, timestamped, outcome-labelled data, which makes it one of the few domains where a forecasting model can be scored honestly and continuously. Every prediction is settled within hours, so a model that is confidently wrong has nowhere to hide.

That property is why we build here. The question is almost never "who wins" — a single predicted scoreline is close to useless. It is "what is the full distribution of outcomes, and is that distribution honest": how often does something we called 70% actually happen.

• Probabilistic modelling of the sport itself, starting from a well-understood baseline rather than throwing a neural network at a table. A bivariate Poisson family fits football scoring; a multiplicative log-linear model fits run scoring in baseball. Anything newer has to beat that baseline on calibration before it ships.
• Feature engineering that is strictly walk-forward, so a match is only ever predicted from data that existed before it. Rolling form windows, ELO, head-to-head history and park factors are all built this way — leakage is the failure mode that makes an offline model look excellent and a live one look random.
• Monte Carlo simulation as the integration layer. Models produce distribution parameters; tens of thousands of simulated matches turn those into the full outcome space — result, totals, both-teams-to-score, handicaps, exact scorelines — and carry the uncertainty through instead of discarding it.
• Calibration measurement as the acceptance test. Brier score, log-loss, ranked probability score and expected calibration error, not accuracy. A model that is right slightly more often but wildly overconfident is worse than a modest one that knows what it does not know.
• Ingestion and scheduling around real season calendars, so datasets refresh and models retrain on their own rather than when someone remembers.

We run Sports Predictor, an internal system covering football and baseball, described on our Engineering page. It is where the approach above is actually implemented, including the parts that did not work: it exists partly because an earlier gradient-boosted regression learned to predict the league average, and the log-linear reformulation that replaced it is the reason the current model distinguishes a strong offence from a weak one at all.