Validated on the strict 2018-2022 window and confirmed on the untouched
2022-2026 test set (RPS 0.1703 vs 0.1721 over 4,448 matches):
- the Elo member now blends 30% of a 3x-faster Elo walk, so recent
results move ratings much harder
- ensemble weight shifts from 75/25 toward Elo to 45/55 toward the
time-decayed Team-DC member — the recent-form model now leads
- Team-DC refits nightly (and at boot) on the 15-year window plus every
finished 2026 match, via a new committed-at-build dcTrain.json
(server-only, excluded from the PWA precache)
- a recent-form goal multiplier was also tested and did NOT validate;
it ships disabled, and backtest.json records the whole decision
buildBacktest.ts grew the experiment harness: wider xi/k grids, the
fast-Elo and form variants, a pre-registered ship bar (>=0.0005
validation RPS), and a 7-day refit-cadence check. Older ratings.json
files still work — all new model fields are optional with golden
regression coverage.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- src/lib/model/teamDc.ts: per-team attack/defence Dixon-Coles via weighted MLE
(Maher iterations), exponential time-decay, shrinkage for thin histories,
learned home multiplier; log-pool of score matrices. 7 vitest cases.
- buildBacktest.ts rewritten as a strict three-way-split bake-off: params
pre-2018, tune (xi, k, w) on 2018-2022 validation, final scores on untouched
2022-2026 test (4,448 matches). Result: ensemble RPS 0.1721 beats Elo-DC
0.1726 and Team-DC 0.1801; ECE stays 0.01. Tuned: xi=0.5 k=5 w=0.75.
- Runtime now IS the backtested model: ratings.json carries teamDc+ensembleW;
matchMatrix() pools both members for predictions, Monte Carlo and advance
probs; ModelEngine re-rating preserves ensemble params; lambdas reported as
matrix expectations (drives in-play).
- scripts/buildSquadValues.ts: Transfermarkt FIWC participants page -> all 48
squad market values (France 1.52bn ... Qatar 20m). Display/availability layer
only — June-2026 values are NOT in the backtested model (would leak).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- scripts/buildBacktest.ts: honest walk-forward validation — params fit on
pre-2018 internationals, tested out-of-sample on 7,988 matches (2018-2026)
predicting each game from prior data only. Proper scoring (Brier/log-loss/RPS/
accuracy) vs uniform / base-rate / Elo-only baselines + a calibration analysis
(reliability bins + ECE). Results: 60% accuracy, RPS 0.171 (beats all
baselines), ECE 0.01 (excellent calibration).
- Methodology page (/methodology): plain-language model walkthrough, the backtest
scorecard, a custom-SVG reliability diagram, and honest limits. Transparency is
the differentiator — no market odds, no overclaiming.
- ReliabilityDiagram component; 'Model' nav entry.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
NilsBriggen