Rugby Moneyball: Win-Probability Tool

The model under the hood is a logistic regression with L2 regularisation, trained on the Kintetsu Liners 2025 season. 13 matches across wins and losses, eight features. Predictions run entirely in your browser; nothing you enter is sent anywhere.

See the methodology essay for what the coefficients mean and what the model can and can’t tell you.

Quick-load · real Kintetsu matches

Match inputs

Yellow cardscount per match

Turnover differentialwon − conceded

Points per carrypoints / carries

Penalties concededcount per match

Line breaks / defenders beatencleanness ratio

Possession %0 to 1

Dominant carry %dominant / total

Missed tackle %0 to 1

Predicted win probabilityPredicted win

69%

baseline 75%

0%100%

Figure // SHAP-style attributionlogit contributions

Each bar = standardised feature value × model coefficient. Bars right of zero push toward win, left of zero toward loss. Sum + intercept → log-odds → probability. Sorted by magnitude.

Model honesty

Trained on 13 matches (Kintetsu Liners 2025). AUC 0.60 with 95% bootstrap CI [0.08, 1.00]. The lower bound sits below 0.5 — meaning on a different season the model could plausibly be useless. The naive “always W” baseline beats it on accuracy. Illustrative, not deployable. See the methodology essay for the full picture.

Pt 1 · The Metrics That Actually Predict Wins Pt 2 · Predicting Wins With Machine LearningPt 3 · This tool (you are here)Pt 4 · Player Recruitment Profiles · comingPt 5 · Does It Generalise? · coming