Research Papers

Introduced the cornerstone bivariate Poisson model for football score prediction, with a time-decay correction and a low-score correction term (the Dixon-Coles adjustment) that significantly improved accuracy for 0-0, 1-0, 0-1, and 1-1 scorelines.

Established the independent Poisson model for football scores, demonstrating that home and away goals can be modelled as independent Poisson random variables with attack/defence strength parameters — the framework Model90's Poisson component is built on.

Extended the Dixon-Coles framework with a Bayesian dynamic model in which team attack/defence parameters evolve over time. Demonstrates significantly better performance than the static model, especially early in the season.

Introduced the Elo rating system originally for chess, subsequently adopted by FIFA for international football. Model90 uses Elo ratings as a core feature capturing each team's relative historical strength and recent form.

Compared Elo-based and Dixon-Coles models for international football prediction, demonstrating that simple Elo variants can outperform more complex models in out-of-sample forecasting when data is limited — a key insight for World Cup predictions.

Introduced pi-ratings, a Bayesian approach for real-time updating of team strength, and evaluated the value of including betting market odds as a feature. Showed that market-implied probabilities carry significant information beyond what statistical models capture alone.

Applied gradient boosting and random forests to football prediction with a rich feature set including form, head-to-head records, squad ratings, and market odds. Demonstrated that ensemble methods consistently outperform logistic regression and Poisson-based approaches on a large held-out test set.

Proposed a two-stage framework combining expected goals (xG) as a performance metric with market-implied probabilities and historical features. Found that xG captures team quality more accurately than goals alone, reducing noise from luck and producing better-calibrated outcome probabilities.