Model catalogue

Models

Latest evaluation: line-move-v1b

2026-07-24T23:07:48+00:00 · commit fa69d83

Every model in the bench with its role, verdict, and latest walk-forward metrics. λ is the market-blend weight (how much information the model carries beyond the market); CLA is closing-line agreement (does the market move toward the model's picks?). Verdicts are review-owned — see the Docs page for the audit trail.

| |

H2H

Candidate H2H

★ ML Lineup

lineup_margin_ridge · LightGBM + team-list features

Walk-forward LightGBM trained on Elo diff, squad-strength deltas, rest, and lineup composition features. Writes to gold_nrl_predictions; shown as the starred row in round-pred reports.

acc61.6% fromcatalogue

Best accuracy on the bench at 61.6% (acc) vs 60.2% market baseline. Used as the primary H2H predictor since 2026-07-10. Refresh with: python ml/predict_weekly_margin.py --season 2026 --round <N>.

Candidate H2H open priced

Logit + RAPM Lineup

logit_rapm_v1 · Logistic regression + RAPM

logit_open_mkt + walk-forward RAPM lineup delta (player plus-minus from team lists, identity-resolved via gold_nrl_player_name_map).

log-loss0.629 λ0.1 CLA0.5651 features16 fromline-v2-final

SIGNAL CONFIRMED, GATE NOT CLEARED (phase3-rapm-v3, real deltas after the join fix): λ=0.141 — first candidate above the 0.10 bar — and CLA z=5.22 (the market moves toward its picks). G6 missed by 0.0005 log-loss vs the ablation control. Decisive test: the team-list-timing channel (rapm_timing_v1 paper trades).

Features (16)

Candidate H2H open priced

State-Space Strength

ss_strength_v1 · Kalman state-space

Time-varying attack/defence strengths with uncertainty; season-boundary variance inflation replaces the Elo regression trick.

log-loss0.6377 λ0.055 CLA0.5442 fromline-v2-final

KEPT (G1 pass — beats Elo on log-loss with the best raw calibration on the bench; G2 fail, λ=0.055). A better Elo, not yet an edge.

Baseline H2H open priced

Elo Current

elo_current · Elo

Margin-scaled Elo baseline (K=24, HFA+55, 25% season regression). The backbone most other features derive from.

log-loss0.6399 λ0.015 CLA0.5187 fromline-v2-final

Incumbent baseline. Log-loss ~0.640; adds no edge over the market (λ≈0.02).

Baseline H2H open priced

Logit Core

logit_core · Logistic regression

Walk-forward logit on Elo + rolling form + schedule + weather. Actionable at open.

log-loss0.6424 λ0.021 CLA0.537 features14 fromline-v2-final

Incumbent. Mostly Elo restated; λ≈0.02.

Features (14)

Challenger H2H

LightGBM v1

lgbm_v1 · Gradient boosting

Offline challenger (nrl/src). Retrained leakage-free July 2026; serves the CLI predictor with --explain (TreeSHAP).

log-loss0.6533 acc61.9%

Diagnostic only — BELOW market baseline since the 2026-07-16 retrain (acc 61.9% vs 62.45% implied-prob baseline). Fixed 2024 train cutoff and pre-backfill weather rows; regenerate training data with a rolling cutoff or retire (decision queued, review-2026-07-16). Must not drive trade decisions.

Control H2H open priced

Logit + Open Market (control)

logit_open_mkt · Logistic regression

Ablation control: logit_core features + opening implied probability, NO lineup delta. G6's incumbent — candidates must beat this, not bare logit_core.

log-loss0.6287 λ0.086 CLA0.5291 features15 fromline-v2-final

Control only — never bet, never promoted. Exists so RAPM passes can only come from the lineup delta itself.

Features (15)

Diagnostic H2H close priced

Lineup Prior Year

lineup_prior_year_v1 · Logistic regression

Closing-market features + prior-year lineup aggregates.

log-loss0.6149 λ0.076 CLA0.7788 features22 fromline-v2-final

Diagnostic only; lineup coefficients historically ~0.

Features (22)

Diagnostic H2H close priced

Logit Core + Market

logit_core_market · Logistic regression

Includes CLOSING market features — measures how much of the close the core features explain.

log-loss0.6141 λ0.052 CLA0.8079 features17 fromline-v2-final

Diagnostic only. Never actionable, never seeded into paper trading.

Features (17)

Line

Candidate Line

Line Move v1

line_move_v1 · Calibrated logistic direction model

Tuesday post-team-list prediction of closing-line direction, placed at an immutable real opening handicap.

metricsnot yet evaluated

LIVE PAPER only. Backtested chosen-side CLV +1.47 points (95% CI +1.08 to +1.87); review at 20/40/100 settled reads.

Candidate Line open priced

Line Ridge Core

line_ridge_core · Ridge margin

Ridge on margin vs the OPENING line; normal-CDF cover probabilities.

MAE13.644 features29 fromline-v2-final

Strongest signal in the shop: +4.4% ROI over ~1,580 open-line bets, CI [−4.4%, +19%] — promising, not significant. Priority for validation.

Features (29)

elo_home_probelo_rating_diffrolling_margin_diff_l5rolling_win_rate_diff_l5rest_days_diffplayoff_flagneutral_venue_flagtemperature_2mweather_condition_binprecipitation_mm_pre_match_3hwind_speed_binis_crosswind_flagweather_data_availableweather_impact_flagabs_elo_rating_diffrolling_home_points_for_l5rolling_away_points_for_l5rolling_home_points_against_l5rolling_away_points_against_l5rolling_points_for_sum_l5rolling_points_against_sum_l5rolling_total_points_l5season_to_date_avg_totalrecent_league_avg_total_l20lineup_overall_deltalineup_halves_deltalineup_edge_deltalineup_availablelineup_nonzero_rate

Candidate Line open priced

Line v2a — registered bundle

line_v2a · Ridge margin + lineup bundle

Research-only walk-forward ridge using the pre-registered lineup bundle; games-together is excluded when VIF exceeds 5.

MAE13.586 features33 fromline-v2-final

Research result: MAE/CLA joint gates not cleared. Never wired to live prediction.

Features (33)

Candidate Line open priced

Line v2a — HistGradientBoosting challenger

line_v2a_hgb · Gradient boosting margin + lineup bundle

Registered research challenger with nested time-ordered parameter selection.

MAE13.611 features33 fromline-v2-final

Research-only challenger; not wired to live prediction.

Features (33)

Candidate Line open priced

Line v2a — Huber challenger

line_v2a_huber · Huber margin + lineup bundle

Registered research challenger with nested time-ordered epsilon/alpha selection.

MAE13.624 features33 fromline-v2-final

Research-only challenger; not wired to live prediction.

Features (33)

Parked Line open priced

Compound-Poisson Margin

cpois_margin_v1 · Try/conversion scoring model

Dixon-Coles-style try rates → full margin distribution.

MAE14.297 fromline-v2-final

PARKED — G4 fail (MAE 14.3 vs ridge 13.5). One more iteration allowed (feed ss_strength states, fit overdispersion), then retire.

Diagnostic Line close priced

Line Ridge + Market

line_ridge_market · Ridge margin

Closing-line features; diagnostic twin of line_ridge_core.

MAE13.421 features38 fromline-v2-final

Diagnostic only.

Features (38)

elo_home_probelo_rating_diffrolling_margin_diff_l5rolling_win_rate_diff_l5rest_days_diffplayoff_flagneutral_venue_flagtemperature_2mweather_condition_binprecipitation_mm_pre_match_3hwind_speed_binis_crosswind_flagweather_data_availableweather_impact_flagclose_home_implied_probopen_home_implied_probmarket_moveabs_elo_rating_diffrolling_home_points_for_l5rolling_away_points_for_l5rolling_home_points_against_l5rolling_away_points_against_l5rolling_points_for_sum_l5rolling_points_against_sum_l5rolling_total_points_l5season_to_date_avg_totalrecent_league_avg_total_l20lineup_overall_deltalineup_halves_deltalineup_edge_deltalineup_availablelineup_nonzero_ratess_margin_varclose_home_lineopen_home_linehome_line_moveclose_home_line_implied_probopen_home_line_implied_prob

Total

Baseline Total open priced

Total Ridge Core

total_ridge_core · Ridge total

Ridge on total points vs the opening total.

MAE10.283 features25 fromline-v2-final

Roughly breakeven at open; incumbent for the totals track.

Features (25)

abs_elo_rating_diffrolling_home_points_for_l5rolling_away_points_for_l5rolling_home_points_against_l5rolling_away_points_against_l5rolling_points_for_sum_l5rolling_points_against_sum_l5rolling_total_points_l5season_to_date_avg_totalrecent_league_avg_total_l20rest_days_diffplayoff_flagneutral_venue_flaglineup_overall_deltalineup_halves_deltalineup_edge_deltalineup_availablelineup_nonzero_ratetemperature_2mweather_condition_binprecipitation_mm_pre_match_3hwind_speed_binis_crosswind_flagweather_data_availableweather_impact_flag

Parked Total open priced

Compound-Poisson Total

cpois_total_v1 · Try/conversion scoring model

Same scoring model, totals track.

MAE10.649 fromline-v2-final

PARKED — G3 fail (MAE 10.6 vs ridge 10.3).

Diagnostic Total close priced

Total Ridge + Market

total_ridge_market · Ridge total

Closing-total features; diagnostic twin.

MAE10.082 features30 fromline-v2-final

Diagnostic only.

Features (30)