The Model

Round 17updated Thu 25 Jun 2026, 06:15AM AEST

A statistical NRL model: try-scorer, goal-kicking and player-points projections priced against live multi-bookmaker odds to surface value. Educational — not betting advice.

8
Matches
966
Priced markets
1136
Pick'em rows
Lineups
Confirmed team lists for every game this round, with each side's goal kicker flagged.
Predictions
Per-player projected tries, points and kicker points for every game this round.
Compare odds
Model fair price next to live Sportsbet, Ladbrokes, TAB, PointsBet & Dabble prices — best highlighted.
Value bets
Every +EV market ranked by edge, with the book offering it and a suggested Kelly stake.
Pick'em
Type any Dabble line and the model returns P(over), fair odds and a lean — build a parlay slip.
Scoring
Player-points and try-scorer leaders with the model's price versus the best book.
Accuracy
Out-of-sample backtest: try-scorer AUC, calibration curve and per-stat error vs a form baseline.

How it works

1Form features

From Champion Data match stats + the confirmed team lists, we build each player’s recent-form features — shifted one game so a prediction never sees the game it’s predicting.

2The models

Gradient-boosted trees predict a try-scorer’s expected tries (Poisson rate), a team’s goal-kicker output (anchored to the named kicker), and stat means for tackles, metres and performance points.

3Distributions

A projection isn’t a bet — we turn it into a distribution: Poisson for tries, a convolution of 4×tries + 2×goals for points, and Normal(mean, σ) for the rest, with σ calibrated from out-of-sample error.

4Fair price

Summing the distribution above/below a line gives the win probability for that market; fair odds = 1 ÷ probability, with no bookmaker margin.

5Find value

Live odds from 5 books are de-vigged to their true probability. EV = model prob × best price − 1; a Kelly stake sizes the bet to the edge.

6Backtested

Every figure is out-of-sample — trained on earlier seasons, scored on 2023–25: try-model AUC ≈ 0.73, calibrated, and 25–45% lower error than a recent-form baseline.

The short version: trees predict each player’s rate or mean → that becomes a probability distribution → distributions price every market line → compared against de-vigged book odds to surface +EV. See the backtest for how it actually performs.

The model runs in a separate pipeline and refreshes through the day. Figures are for information and entertainment only — not betting advice. Gamble responsibly.