Rugby Moneyball: win probability and player profile tool
The match model is LASSO logistic regression with class-balanced weighting, trained on 302 team-match outcomes pooled across Japan Rugby League One D1 23/24 (n=194) + D2 24/25 (n=108), with 13 inputs. Pooled leave-one-round-out AUC 0.894 (95% bootstrap CI [0.86, 0.93]); cross-league transfer 0.86–0.89.
The player model is the Pt 4 Layer 2 fit: position-controlled OLS on 34 Bronco-tested Kintetsu players (median 7 player-matches each across both seasons). Only height has a bootstrap-stable effect on the carry-metres family; other physical predictors come with wide uncertainty. The profile output is illustrative and the disclosure says so on the predict pane.
Both predictions run in the browser; nothing you enter is sent anywhere. URL state syncs for the match tab so the prediction is shareable.
LASSO logistic regression, trained on 302 team-matches from Japan Rugby League One D1 23/24 (n=194) + D2 24/25 (n=108). 13 inputs, 3 of which the LASSO sets to zero weight. Pooled leave-one-round-out AUC 0.894, 95% bootstrap CI [0.86, 0.93]. The 302 rows are 151 fixtures seen from both sides, so they are not independent; resampling whole fixtures widens the interval by about fifteen percent. Single-team predictions still vary widely. Strong illustrative model, not a tactical decision tool. The methodology essays are not yet published.