v3 Phase B core: Team-DC + backtested ensemble, squad values

- 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>
This commit is contained in:
2026-06-11 17:23:51 +02:00
parent d887664fce
commit f3b2a69b31
9 changed files with 549 additions and 122 deletions
+1
View File
@@ -0,0 +1 @@
{"fetchedAt":"2026-06-11T15:14:56.650Z","source":"transfermarkt.com","teams":{"France":{"totalValue":1520000000,"avgValue":58580000},"England":{"totalValue":1360000000,"avgValue":52430000},"Spain":{"totalValue":1220000000,"avgValue":47030000},"Portugal":{"totalValue":1010000000,"avgValue":38670000},"Germany":{"totalValue":947000000,"avgValue":36420000},"Brazil":{"totalValue":928200000,"avgValue":35700000},"Argentina":{"totalValue":782500000,"avgValue":31300000},"Netherlands":{"totalValue":754200000,"avgValue":29010000},"Norway":{"totalValue":589900000,"avgValue":22690000},"Belgium":{"totalValue":547500000,"avgValue":21060000},"Ivory Coast":{"totalValue":522100000,"avgValue":20080000},"Senegal":{"totalValue":478100000,"avgValue":18390000},"Turkey":{"totalValue":473700000,"avgValue":18220000},"Morocco":{"totalValue":447700000,"avgValue":17220000},"Sweden":{"totalValue":406080000,"avgValue":15620000},"Croatia":{"totalValue":387300000,"avgValue":14900000},"USA":{"totalValue":385650000,"avgValue":14830000},"Ecuador":{"totalValue":368700000,"avgValue":14180000},"Uruguay":{"totalValue":359300000,"avgValue":13820000},"Switzerland":{"totalValue":332500000,"avgValue":12790000},"Colombia":{"totalValue":302350000,"avgValue":11630000},"Japan":{"totalValue":270850000,"avgValue":10420000},"Algeria":{"totalValue":256900000,"avgValue":9880000},"Austria":{"totalValue":245200000,"avgValue":9430000},"Ghana":{"totalValue":234600000,"avgValue":8690000},"Canada":{"totalValue":198650000,"avgValue":7640000},"Mexico":{"totalValue":191850000,"avgValue":7380000},"Czech Republic":{"totalValue":188180000,"avgValue":7240000},"Scotland":{"totalValue":170250000,"avgValue":6550000},"Paraguay":{"totalValue":153650000,"avgValue":5910000},"Bosnia & Herzegovina":{"totalValue":151600000,"avgValue":5830000},"DR Congo":{"totalValue":143900000,"avgValue":5530000},"South Korea":{"totalValue":139050000,"avgValue":5350000},"Egypt":{"totalValue":116480000,"avgValue":4480000},"Uzbekistan":{"totalValue":85330000,"avgValue":3280000},"Australia":{"totalValue":77450000,"avgValue":2980000},"Tunisia":{"totalValue":69950000,"avgValue":2690000},"Haiti":{"totalValue":55900000,"avgValue":2150000},"Cape Verde":{"totalValue":54500000,"avgValue":2100000},"South Africa":{"totalValue":49250000,"avgValue":1890000},"Saudi Arabia":{"totalValue":40680000,"avgValue":1560000},"Panama":{"totalValue":34550000,"avgValue":1330000},"New Zealand":{"totalValue":34350000,"avgValue":1320000},"Iran":{"totalValue":32050000,"avgValue":1230000},"Curaçao":{"totalValue":25780000,"avgValue":991000},"Iraq":{"totalValue":21200000,"avgValue":815000},"Jordan":{"totalValue":20300000,"avgValue":781000},"Qatar":{"totalValue":19930000,"avgValue":766000}}}
+169 -108
View File
@@ -1,14 +1,16 @@
// Walk-forward backtest of the Elo + Dixon-Coles model on historical
// internationals, with an honest train/test split (params fit on pre-2018,
// tested out-of-sample on 2018→now). Outputs proper scoring metrics (Brier,
// log-loss, RPS, accuracy), baseline comparisons, and a calibration/reliability
// analysis → public/data/backtest.json, shown on the Methodology page.
// v3 bake-off backtest with a strict three-way split — the evidence behind the
// shipped model. NO information leaks forward:
// params (pre-2018) : Elo→goals calibration, Dixon-Coles rho
// validate (20182022) : tune Team-DC decay ξ, shrinkage k, ensemble weight w
// test (20222026) : untouched final numbers for every variant + baselines
// Output → public/data/backtest.json (Methodology page).
import { readFileSync, writeFileSync, mkdirSync } from 'node:fs';
import { fileURLToPath } from 'node:url';
import { dirname, join } from 'node:path';
import { canonicalTeam } from '../src/lib/teams';
import { HOME_ADV_ELO, importanceWeight, eloDelta } from '../src/lib/model/elo';
import { lambdasFromElo, scoreMatrix, outcomeProbs, poissonPmf, type ModelParams } from '../src/lib/model/poisson';
import { fitTeamDc, teamDcLambdas, poolMatrices, type DcMatch, type TeamDcParams } from '../src/lib/model/teamDc';
const ROOT = join(dirname(fileURLToPath(import.meta.url)), '..');
const CSV = join(ROOT, 'data', 'raw', 'international-results.csv');
@@ -16,13 +18,21 @@ const OUT_DIR = join(ROOT, 'public', 'data');
const START_ELO = 1500;
const PARAM_FROM = '2006-01-01';
const TRAIN_END = '2018-01-01'; // params fit on [PARAM_FROM, TRAIN_END)
const TEST_TO = '2026-06-01'; // exclude the 2026 World Cup itself
const VAL_FROM = '2018-01-01';
const TEST_FROM = '2022-01-01';
const TEST_TO = '2026-06-01';
const REFIT_DAYS = 60;
const FIT_WINDOW_YEARS = 15;
interface Row { date: string; home: string; away: string; hs: number; as: number; tournament: string; neutral: boolean }
const XI_GRID = [0.5, 1, 1.5, 2, 2.5];
const K_GRID = [5, 10, 20];
interface Row { date: string; home: string; away: string; hs: number; as: number; tournament: string; neutral: boolean; t: number }
type Probs = { h: number; d: number; a: number };
type Outcome = 'h' | 'd' | 'a';
const DAY = 24 * 60 * 60 * 1000;
function parse(): Row[] {
const lines = readFileSync(CSV, 'utf8').split('\n');
const rows: Row[] = [];
@@ -31,9 +41,13 @@ function parse(): Row[] {
if (!c || c.length < 9) continue;
const hs = Number(c[3]); const as = Number(c[4]);
if (!Number.isFinite(hs) || !Number.isFinite(as)) continue;
rows.push({ date: c[0]!, home: c[1]!, away: c[2]!, hs, as, tournament: c[5]!, neutral: c[8]!.trim().toUpperCase() === 'TRUE' });
rows.push({
date: c[0]!, home: canonicalTeam(c[1]!), away: canonicalTeam(c[2]!), hs, as,
tournament: c[5]!, neutral: c[8]!.trim().toUpperCase() === 'TRUE',
t: new Date(c[0]!).getTime(),
});
}
rows.sort((a, b) => a.date.localeCompare(b.date));
rows.sort((a, b) => a.t - b.t);
return rows;
}
@@ -46,120 +60,165 @@ function dcTau(h: number, a: number, lh: number, la: number, rho: number): numbe
}
const outcomeOf = (hs: number, as: number): Outcome => (hs > as ? 'h' : hs < as ? 'a' : 'd');
const probsOf = (m: number[][]): Probs => { const p = outcomeProbs(m); return { h: p.home, d: p.draw, a: p.away }; };
// ---- scoring ----
function brier(p: Probs, o: Outcome): number {
return (p.h - (o === 'h' ? 1 : 0)) ** 2 + (p.d - (o === 'd' ? 1 : 0)) ** 2 + (p.a - (o === 'a' ? 1 : 0)) ** 2;
}
function logloss(p: Probs, o: Outcome): number {
return -Math.log(Math.max(1e-12, p[o]));
}
const logloss = (p: Probs, o: Outcome): number => -Math.log(Math.max(1e-12, p[o]));
function rps(p: Probs, o: Outcome): number {
// ordered H, D, A
const y = { h: o === 'h' ? 1 : 0, d: o === 'd' ? 1 : 0, a: o === 'a' ? 1 : 0 };
const c1p = p.h, c1y = y.h;
const c2p = p.h + p.d, c2y = y.h + y.d;
return ((c1p - c1y) ** 2 + (c2p - c2y) ** 2) / 2;
const y1 = o === 'h' ? 1 : 0;
const y2 = o === 'a' ? 0 : 1;
return ((p.h - y1) ** 2 + (p.h + p.d - y2) ** 2) / 2;
}
const argmax = (p: Probs): Outcome => (p.h >= p.d && p.h >= p.a ? 'h' : p.d >= p.a ? 'd' : 'a');
function scoreSet(preds: { p: Probs; o: Outcome }[]) {
let b = 0, l = 0, r = 0, correct = 0;
for (const { p, o } of preds) { b += brier(p, o); l += logloss(p, o); r += rps(p, o); if (argmax(p) === o) correct++; }
const n = preds.length;
return { brier: b / n, logloss: l / n, rps: r / n, accuracy: correct / n };
const n = Math.max(1, preds.length);
return {
brier: +(b / n).toFixed(4), logloss: +(l / n).toFixed(4),
rps: +(r / n).toFixed(4), accuracy: +(correct / n).toFixed(4),
};
}
function main(): void {
const rows = parse();
const tValFrom = new Date(VAL_FROM).getTime();
const tTestFrom = new Date(TEST_FROM).getTime();
const tTestTo = new Date(TEST_TO).getTime();
// ---------- pass 1: Elo walk + Elo-DC param fit on pre-2018 ----------
const elo = new Map<string, number>();
const get = (t: string) => elo.get(t) ?? START_ELO;
// ---- fit params on the training window ----
let sumTotal = 0, nCal = 0, sxy = 0, sxx = 0;
const calib: { d: number; h: number; a: number }[] = [];
// we need a first pass to fit params using pre-match Elo; do it inline while walking
// (Elo is updated every match; calibration accumulates only within the train window)
/** pre-match Elo per row index (so later passes never see post-match info) */
const preElo: { eh: number; ea: number }[] = new Array(rows.length);
const testPreds: { p: Probs; o: Outcome }[] = [];
const baseUniform: { p: Probs; o: Outcome }[] = [];
const baseRate: { p: Probs; o: Outcome }[] = [];
const baseElo: { p: Probs; o: Outcome }[] = [];
for (let i = 0; i < rows.length; i++) {
const r = rows[i]!;
const eh = get(r.home), ea = get(r.away);
preElo[i] = { eh, ea };
const homeAdv = r.neutral ? 0 : HOME_ADV_ELO;
if (r.date >= PARAM_FROM && r.t < tValFrom) {
const eff = eh - ea + homeAdv;
sumTotal += r.hs + r.as; nCal++; sxy += eff * (r.hs - r.as); sxx += eff * eff;
calib.push({ d: eff, h: r.hs, a: r.as });
}
const d = eloDelta(r.hs, r.as, eh, ea, importanceWeight(r.tournament), homeAdv);
elo.set(r.home, eh + d); elo.set(r.away, ea - d);
}
const avgGoals = sumTotal / Math.max(1, nCal);
const goalsPerElo = sxy / Math.max(1e-9, sxx);
let rho = -0.05, bestLL = -Infinity;
for (let r0 = -0.2; r0 <= 0.1 + 1e-9; r0 += 0.02) {
let ll = 0;
for (const m of calib) {
const sup = goalsPerElo * m.d;
const lh = Math.min(8, Math.max(0.15, avgGoals / 2 + sup / 2));
const la = Math.min(8, Math.max(0.15, avgGoals / 2 - sup / 2));
ll += Math.log(Math.max(1e-12, poissonPmf(lh, m.h) * poissonPmf(la, m.a) * dcTau(m.h, m.a, lh, la, r0)));
}
if (ll > bestLL) { bestLL = ll; rho = +r0.toFixed(2); }
}
const params: ModelParams = { goalsPerElo, avgGoals, rho, homeAdvElo: HOME_ADV_ELO };
const eloMatrix = (i: number): number[][] => {
const r = rows[i]!;
const { lambdaHome, lambdaAway } = lambdasFromElo(preElo[i]!.eh, preElo[i]!.ea, params, r.neutral ? 0 : HOME_ADV_ELO);
return scoreMatrix(lambdaHome, lambdaAway, rho);
};
// ---------- walk-forward Team-DC over a window, collecting matrices ----------
type Pred = { i: number; mDc: number[][] };
const walkTeamDc = (xi: number, k: number, fromT: number, toT: number): Pred[] => {
const preds: Pred[] = [];
let fit: TeamDcParams | null = null;
let fitAt = -Infinity;
const windowMs = FIT_WINDOW_YEARS * 365 * DAY;
for (let i = 0; i < rows.length; i++) {
const r = rows[i]!;
if (r.t < fromT) continue;
if (r.t >= toT) break;
if (r.t - fitAt > REFIT_DAYS * DAY) {
const train: DcMatch[] = [];
for (let j = 0; j < i; j++) {
const m = rows[j]!;
if (r.t - m.t > windowMs) continue;
train.push({
daysAgo: (r.t - m.t) / DAY, home: m.home, away: m.away,
homeGoals: m.hs, awayGoals: m.as, neutral: m.neutral,
});
}
fit = fitTeamDc(train, xi, k);
fitAt = r.t;
}
const { lambdaHome, lambdaAway } = teamDcLambdas(fit!, r.home, r.away, r.neutral);
preds.push({ i, mDc: scoreMatrix(lambdaHome, lambdaAway, rho) });
}
return preds;
};
// ---------- validation: tune ξ, k, then w ----------
console.log('tuning on validation (20182022)…');
let best = { xi: 1, k: 10, w: 0.5, rps: Infinity };
for (const xi of XI_GRID) {
for (const k of K_GRID) {
const preds = walkTeamDc(xi, k, tValFrom, tTestFrom);
// pure Team-DC score (w=0 candidate) and weight sweep against Elo
for (let w = 0; w <= 1.0001; w += 0.05) {
let r = 0;
for (const p of preds) {
const pooled = w === 0 ? p.mDc : w === 1 ? eloMatrix(p.i) : poolMatrices(eloMatrix(p.i), p.mDc, w);
r += rps(probsOf(pooled), outcomeOf(rows[p.i]!.hs, rows[p.i]!.as));
}
r /= preds.length;
if (r < best.rps) best = { xi, k, w: +w.toFixed(2), rps: r };
}
console.log(` ξ=${xi} k=${k} done (best so far: ξ=${best.xi} k=${best.k} w=${best.w} rps=${best.rps.toFixed(4)})`);
}
}
// ---------- test window: final scores for every variant ----------
console.log(`testing on ${TEST_FROM}${TEST_TO} with ξ=${best.xi} k=${best.k} w=${best.w}`);
const testPreds = walkTeamDc(best.xi, best.k, tTestFrom, tTestTo);
const setEnsemble: { p: Probs; o: Outcome }[] = [];
const setElo: { p: Probs; o: Outcome }[] = [];
const setDc: { p: Probs; o: Outcome }[] = [];
const setUniform: { p: Probs; o: Outcome }[] = [];
const setRate: { p: Probs; o: Outcome }[] = [];
const reliabilityRaw: { p: number; y: number }[] = [];
let wcCorrect = 0, wcN = 0;
// outcome base rates over the test window (computed in a quick pre-scan)
// outcome base rates over the test window
let th = 0, td = 0, ta = 0, tn = 0;
for (const r of rows) {
if (r.date < TRAIN_END || r.date >= TEST_TO) continue;
const o = outcomeOf(r.hs, r.as); if (o === 'h') th++; else if (o === 'd') td++; else ta++; tn++;
for (const p of testPreds) {
const o = outcomeOf(rows[p.i]!.hs, rows[p.i]!.as);
if (o === 'h') th++; else if (o === 'd') td++; else ta++; tn++;
}
const rate: Probs = { h: th / tn, d: td / tn, a: ta / tn };
// params get finalized at TRAIN_END; predictions before that use a rolling fit.
let params: ModelParams = { goalsPerElo: 0.0057, avgGoals: 2.7, rho: -0.05, homeAdvElo: HOME_ADV_ELO };
let paramsFixed = false;
const finalizeParams = (): ModelParams => {
const avgGoals = sumTotal / Math.max(1, nCal);
const goalsPerElo = sxy / Math.max(1e-9, sxx);
let bestRho = -0.05, bestLL = -Infinity;
for (let rho = -0.2; rho <= 0.1 + 1e-9; rho += 0.02) {
let ll = 0;
for (const m of calib) {
const sup = goalsPerElo * m.d;
const lh = Math.min(8, Math.max(0.15, avgGoals / 2 + sup / 2));
const la = Math.min(8, Math.max(0.15, avgGoals / 2 - sup / 2));
ll += Math.log(Math.max(1e-12, poissonPmf(lh, m.h) * poissonPmf(la, m.a) * dcTau(m.h, m.a, lh, la, rho)));
}
if (ll > bestLL) { bestLL = ll; bestRho = rho; }
}
return { goalsPerElo, avgGoals, rho: Math.round(bestRho * 100) / 100, homeAdvElo: HOME_ADV_ELO };
};
const predict = (eh: number, ea: number, homeAdv: number): Probs => {
const { lambdaHome, lambdaAway } = lambdasFromElo(eh, ea, params, homeAdv);
const p = outcomeProbs(scoreMatrix(lambdaHome, lambdaAway, params.rho));
return { h: p.home, d: p.draw, a: p.away };
};
for (const r of rows) {
if (!paramsFixed && r.date >= TRAIN_END) { params = finalizeParams(); paramsFixed = true; }
const home = canonicalTeam(r.home), away = canonicalTeam(r.away);
const eh = get(home), ea = get(away);
const homeAdv = r.neutral ? 0 : HOME_ADV_ELO;
for (const p of testPreds) {
const r = rows[p.i]!;
const o = outcomeOf(r.hs, r.as);
// accumulate calibration only in the param-fitting window
if (r.date >= PARAM_FROM && r.date < TRAIN_END) {
const effDiff = eh - ea + homeAdv;
sumTotal += r.hs + r.as; nCal++; sxy += effDiff * (r.hs - r.as); sxx += effDiff * effDiff;
calib.push({ d: effDiff, h: r.hs, a: r.as });
const mE = eloMatrix(p.i);
const pooled = poolMatrices(mE, p.mDc, best.w);
const pe = probsOf(mE), pd = probsOf(p.mDc), pp = probsOf(pooled);
setElo.push({ p: pe, o });
setDc.push({ p: pd, o });
setEnsemble.push({ p: pp, o });
setUniform.push({ p: { h: 1 / 3, d: 1 / 3, a: 1 / 3 }, o });
setRate.push({ p: rate, o });
reliabilityRaw.push({ p: pp.h, y: o === 'h' ? 1 : 0 }, { p: pp.d, y: o === 'd' ? 1 : 0 }, { p: pp.a, y: o === 'a' ? 1 : 0 });
if (r.tournament.includes('FIFA World Cup') && !r.tournament.includes('qualification')) {
wcN++; if (argmax(pp) === o) wcCorrect++;
}
// score out-of-sample
if (r.date >= TRAIN_END && r.date < TEST_TO) {
const p = predict(eh, ea, homeAdv);
testPreds.push({ p, o });
baseUniform.push({ p: { h: 1 / 3, d: 1 / 3, a: 1 / 3 }, o });
baseRate.push({ p: rate, o });
// elo-only: expected score → W/A split, fixed draw rate
const e = 1 / (1 + 10 ** ((ea - eh - homeAdv) / 400));
baseElo.push({ p: { h: e * (1 - rate.d), d: rate.d, a: (1 - e) * (1 - rate.d) }, o });
reliabilityRaw.push({ p: p.h, y: o === 'h' ? 1 : 0 }, { p: p.d, y: o === 'd' ? 1 : 0 }, { p: p.a, y: o === 'a' ? 1 : 0 });
if (r.tournament.includes('FIFA World Cup') && !r.tournament.includes('qualification')) {
wcN++; if (argmax(p) === o) wcCorrect++;
}
}
// update Elo (always)
const k = importanceWeight(r.tournament);
const d = eloDelta(r.hs, r.as, eh, ea, k, homeAdv);
elo.set(home, eh + d); elo.set(away, ea - d);
}
// reliability bins (10) + ECE
const bins = Array.from({ length: 10 }, () => ({ sp: 0, sy: 0, n: 0 }));
for (const { p, y } of reliabilityRaw) {
const idx = Math.min(9, Math.floor(p * 10));
@@ -167,33 +226,35 @@ function main(): void {
}
const reliability = bins.map((b) => ({ predicted: b.n ? b.sp / b.n : 0, observed: b.n ? b.sy / b.n : 0, count: b.n }));
const totalN = reliabilityRaw.length;
const ece = bins.reduce((s, b) => s + (b.n ? (b.n / totalN) * Math.abs(b.sp / b.n - b.sy / b.n) : 0), 0);
const ece = +bins.reduce((s, b) => s + (b.n ? (b.n / totalN) * Math.abs(b.sp / b.n - b.sy / b.n) : 0), 0).toFixed(3);
const out = {
generatedAt: new Date().toISOString(),
paramFrom: PARAM_FROM,
trainEnd: TRAIN_END,
trainEnd: VAL_FROM,
validation: { from: VAL_FROM, to: TEST_FROM, tuned: { xi: best.xi, shrinkK: best.k, ensembleW: best.w } },
testFrom: TEST_FROM,
testTo: TEST_TO,
tested: testPreds.length,
params,
model: scoreSet(testPreds),
baselines: {
uniform: scoreSet(baseUniform),
baseRate: scoreSet(baseRate),
eloOnly: scoreSet(baseElo),
},
model: scoreSet(setEnsemble), // the SHIPPED model = ensemble
variants: [
{ name: 'Ensemble (shipped)', ...scoreSet(setEnsemble) },
{ name: 'EloDixon-Coles', ...scoreSet(setElo) },
{ name: 'Team Dixon-Coles', ...scoreSet(setDc) },
],
baselines: { uniform: scoreSet(setUniform), baseRate: scoreSet(setRate), eloOnly: scoreSet(setElo) },
reliability,
ece: Math.round(ece * 1000) / 1000,
worldCup: { tested: wcN, accuracy: wcN ? wcCorrect / wcN : 0 },
ece,
worldCup: { tested: wcN, accuracy: wcN ? +(wcCorrect / wcN).toFixed(4) : 0 },
};
mkdirSync(OUT_DIR, { recursive: true });
writeFileSync(join(OUT_DIR, 'backtest.json'), JSON.stringify(out));
console.log(`backtest → public/data/backtest.json (${out.tested} test matches, ${TRAIN_END}${TEST_TO})`);
console.log(` model: brier ${out.model.brier.toFixed(3)} logloss ${out.model.logloss.toFixed(3)} rps ${out.model.rps.toFixed(3)} acc ${(out.model.accuracy * 100).toFixed(1)}%`);
console.log(` uniform: brier ${out.baselines.uniform.brier.toFixed(3)} logloss ${out.baselines.uniform.logloss.toFixed(3)} rps ${out.baselines.uniform.rps.toFixed(3)}`);
console.log(` baseRate:brier ${out.baselines.baseRate.brier.toFixed(3)} rps ${out.baselines.baseRate.rps.toFixed(3)} | eloOnly rps ${out.baselines.eloOnly.rps.toFixed(3)}`);
console.log(` ECE ${out.ece} | WC accuracy ${(out.worldCup.accuracy * 100).toFixed(1)}% (${out.worldCup.tested})`);
console.log(`\nbacktest → public/data/backtest.json (${out.tested} test matches, ${TEST_FROM}${TEST_TO})`);
for (const v of out.variants) console.log(` ${v.name.padEnd(20)} rps ${v.rps} brier ${v.brier} logloss ${v.logloss} acc ${(v.accuracy * 100).toFixed(1)}%`);
console.log(` baselines: uniform rps ${out.baselines.uniform.rps} | baseRate rps ${out.baselines.baseRate.rps}`);
console.log(` ECE ${ece} | WC accuracy ${(out.worldCup.accuracy * 100).toFixed(1)}% (${wcN})`);
}
main();
+31
View File
@@ -9,6 +9,14 @@ import { dirname, join } from 'node:path';
import { canonicalTeam, ALL_TEAMS } from '../src/lib/teams';
import { HOME_ADV_ELO, importanceWeight, eloDelta, expectedHome } from '../src/lib/model/elo';
import { poissonPmf } from '../src/lib/model/poisson';
import { fitTeamDc, type DcMatch } from '../src/lib/model/teamDc';
// Ensemble hyper-params — tuned on the 20182022 validation window and verified
// out-of-sample on 20222026 by scripts/buildBacktest.ts. Keep in sync with it.
const TEAMDC_XI = 0.5;
const TEAMDC_SHRINK_K = 5;
const ENSEMBLE_W = 0.75; // log-pool weight on the Elo-DC member
const FIT_WINDOW_YEARS = 15;
const ROOT = join(dirname(fileURLToPath(import.meta.url)), '..');
const CSV = join(ROOT, 'data', 'raw', 'international-results.csv');
@@ -119,6 +127,27 @@ function main(): void {
// Ensure every WC team has a rating (default for any never seen).
for (const t of ALL_TEAMS) if (!(t in ratings)) ratings[t] = START_ELO;
// ---- Team-DC ensemble member: fit on the recent window through today ----
const nowT = new Date(lastDate).getTime();
const windowMs = FIT_WINDOW_YEARS * 365 * 24 * 60 * 60 * 1000;
const dcTrain: DcMatch[] = [];
for (const r of rows) {
const t = new Date(r.date).getTime();
if (nowT - t > windowMs) continue;
dcTrain.push({
daysAgo: (nowT - t) / (24 * 60 * 60 * 1000),
home: canonicalTeam(r.home), away: canonicalTeam(r.away),
homeGoals: r.hs, awayGoals: r.as, neutral: r.neutral,
});
}
const teamDc = fitTeamDc(dcTrain, TEAMDC_XI, TEAMDC_SHRINK_K);
// Trim to relevant nations: WC qualifiers + any team rated (keeps file small).
const keep = new Set<string>([...ALL_TEAMS, ...Object.keys(ratings)]);
teamDc.attack = Object.fromEntries(Object.entries(teamDc.attack).filter(([t]) => keep.has(t)).map(([t, v]) => [t, Math.round(v * 1e4) / 1e4]));
teamDc.defence = Object.fromEntries(Object.entries(teamDc.defence).filter(([t]) => keep.has(t)).map(([t, v]) => [t, Math.round(v * 1e4) / 1e4]));
teamDc.mu = Math.round(teamDc.mu * 1e4) / 1e4;
teamDc.homeMult = Math.round(teamDc.homeMult * 1e4) / 1e4;
const out = {
generatedAt: new Date().toISOString(),
asOf: lastDate,
@@ -131,6 +160,8 @@ function main(): void {
rho,
homeAdvElo: HOME_ADV_ELO,
},
teamDc,
ensembleW: ENSEMBLE_W,
ratings,
};
+58
View File
@@ -0,0 +1,58 @@
// Transfermarkt FIWC participants page → public/data/squadvalues.json
// One page lists all 48 squads with total + average market value — far more
// robust than scraping 48 squad pages. Values move slowly, so this is a
// build-time artifact (re-run data:build to refresh). Used by the squad-value
// model layer and the team/squad UI; never overrides results-based ratings.
import { writeFileSync, mkdirSync } from 'node:fs';
import { fileURLToPath } from 'node:url';
import { dirname, join } from 'node:path';
import { canonicalTeam, isKnownTeam } from '../src/lib/teams';
const URL = 'https://www.transfermarkt.com/weltmeisterschaft/teilnehmer/pokalwettbewerb/FIWC';
const OUT_DIR = join(dirname(fileURLToPath(import.meta.url)), '..', 'public', 'data');
function euros(num: string, unit: string): number {
const n = Number(num);
return Math.round(n * (unit === 'bn' ? 1e9 : unit === 'm' ? 1e6 : 1e3));
}
async function main(): Promise<void> {
const res = await fetch(URL, {
headers: {
'User-Agent': 'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/124.0.0.0 Safari/537.36',
'Accept-Language': 'en-US,en;q=0.9',
},
signal: AbortSignal.timeout(30_000),
});
if (!res.ok) throw new Error(`transfermarkt ${res.status}`);
const html = await res.text();
const out: Record<string, { totalValue: number; avgValue: number }> = {};
for (const tr of html.split('<tr').slice(1)) {
const link = tr.match(/<a[^>]*title="([^"]+)"[^>]*href="\/[a-z0-9\-]+\/startseite\/verein\/\d+"/);
const vals = [...tr.matchAll(/€([\d.]+)(bn|m|k)/g)].map((m) => euros(m[1]!, m[2]!));
if (!link || vals.length < 1) continue;
const team = canonicalTeam(link[1]!);
if (!isKnownTeam(team) || out[team]) continue; // 48 qualifiers only, first table wins
const totalValue = Math.max(...vals);
const avgValue = vals.length > 1 ? Math.min(...vals) : Math.round(totalValue / 26);
out[team] = { totalValue, avgValue };
}
const teams = Object.keys(out);
if (teams.length < 40) {
throw new Error(`only parsed ${teams.length}/48 squads — page layout may have changed; aborting (stale file kept)`);
}
mkdirSync(OUT_DIR, { recursive: true });
writeFileSync(
join(OUT_DIR, 'squadvalues.json'),
JSON.stringify({ fetchedAt: new Date().toISOString(), source: 'transfermarkt.com', teams: out }),
);
const top = teams.sort((a, b) => out[b]!.totalValue - out[a]!.totalValue).slice(0, 6)
.map((t) => `${t}${(out[t]!.totalValue / 1e9).toFixed(2)}bn`).join(', ');
console.log(`squad values → public/data/squadvalues.json (${teams.length} teams)`);
console.log(` top: ${top}`);
}
main().catch((e) => { console.error(e); process.exit(1); });
Binary file not shown.
+4 -8
View File
@@ -1,7 +1,7 @@
import type { Fixture, TeamOdds } from '../types';
import { rankGroup, type Result } from '../standings';
import { lambdasFromElo, scoreMatrix, outcomeProbs, type OutcomeProbs } from './poisson';
import { knockoutAdvanceProb, ratingOf, type RatingsModel } from './predict';
import { outcomeProbs, type OutcomeProbs } from './poisson';
import { knockoutAdvanceProb, matchMatrix, ratingOf, type RatingsModel } from './predict';
import { hostAdvantage } from './hosts';
export type { TeamOdds };
@@ -60,10 +60,7 @@ function buildScoreCdf(
home: string, away: string, venue: string, model: RatingsModel,
): { h: number; a: number; cum: number }[] {
const homeAdv = hostAdvantage(home, away, venue, model.params.homeAdvElo);
const { lambdaHome, lambdaAway } = lambdasFromElo(
ratingOf(model, home), ratingOf(model, away), model.params, homeAdv,
);
const m = scoreMatrix(lambdaHome, lambdaAway, model.params.rho);
const m = matchMatrix(home, away, model, homeAdv); // ensemble when configured
const cdf: { h: number; a: number; cum: number }[] = [];
let cum = 0;
for (let h = 0; h < m.length; h++) {
@@ -147,8 +144,7 @@ export function buildSimulator(fixtures: Fixture[], model: RatingsModel) {
let p = advMemo.get(key);
if (p === undefined) {
const homeAdv = hostAdvantage(home, away, venue, model.params.homeAdvElo);
const { lambdaHome, lambdaAway } = lambdasFromElo(ratingOf(model, home), ratingOf(model, away), model.params, homeAdv);
const probs: OutcomeProbs = outcomeProbs(scoreMatrix(lambdaHome, lambdaAway, model.params.rho));
const probs: OutcomeProbs = outcomeProbs(matchMatrix(home, away, model, homeAdv));
p = knockoutAdvanceProb(probs, ratingOf(model, home), ratingOf(model, away));
advMemo.set(key, p);
}
+54 -6
View File
@@ -7,12 +7,18 @@ import {
type OutcomeProbs,
type Scoreline,
} from './poisson';
import { teamDcLambdas, poolMatrices, type TeamDcParams } from './teamDc';
/** The ratings.json payload (also the runtime model state after live re-rating). */
export interface RatingsModel {
asOf: string;
params: ModelParams;
ratings: Record<string, number>;
/** Team-level Dixon-Coles strengths (second ensemble member); optional so a
* v2 ratings file still works as pure Elo-DC. */
teamDc?: TeamDcParams;
/** Log-pool weight on the Elo-DC member (backtest-tuned). */
ensembleW?: number;
}
export const DEFAULT_ELO = 1500;
@@ -33,6 +39,50 @@ export interface MatchPrediction {
scorelines: Scoreline[];
}
/**
* The full scoreline distribution for a match: Elo-DC matrix, log-pooled with
* the Team-DC matrix when the model carries ensemble params (the backtested
* shipped configuration). `homeAdv` is the SIGNED Elo bump for the home slot:
* positive = home side is the host at home, negative = the away side is, 0 = neutral.
*/
export function matchMatrix(
home: string,
away: string,
model: RatingsModel,
homeAdv = 0,
): number[][] {
const { lambdaHome, lambdaAway } = lambdasFromElo(
ratingOf(model, home), ratingOf(model, away), model.params, homeAdv,
);
const mElo = scoreMatrix(lambdaHome, lambdaAway, model.params.rho);
if (!model.teamDc || model.ensembleW == null || model.ensembleW >= 1) return mElo;
// Team-DC orientation: homeAdv > 0 → home side at home; < 0 → away side at
// home (compute flipped, then swap); 0 → neutral.
let dcHome: number, dcAway: number;
if (homeAdv < 0) {
const f = teamDcLambdas(model.teamDc, away, home, false);
dcHome = f.lambdaAway; dcAway = f.lambdaHome;
} else {
const f = teamDcLambdas(model.teamDc, home, away, homeAdv === 0);
dcHome = f.lambdaHome; dcAway = f.lambdaAway;
}
const mDc = scoreMatrix(dcHome, dcAway, model.params.rho);
return poolMatrices(mElo, mDc, model.ensembleW);
}
/** Expected goals per side under a scoreline matrix. */
function matrixLambdas(m: number[][]): { lambdaHome: number; lambdaAway: number } {
let lh = 0, la = 0;
for (let i = 0; i < m.length; i++) {
for (let j = 0; j < m[i]!.length; j++) {
lh += i * m[i]![j]!;
la += j * m[i]![j]!;
}
}
return { lambdaHome: lh, lambdaAway: la };
}
/**
* Win/draw/loss + likely scores for one match. `homeAdv` is the Elo bump for the
* home slot (0 neutral; use hostAdvantage() for host nations at home).
@@ -43,15 +93,13 @@ export function predictMatch(
model: RatingsModel,
homeAdv = 0,
): MatchPrediction {
const eloHome = ratingOf(model, home);
const eloAway = ratingOf(model, away);
const { lambdaHome, lambdaAway } = lambdasFromElo(eloHome, eloAway, model.params, homeAdv);
const m = scoreMatrix(lambdaHome, lambdaAway, model.params.rho);
const m = matchMatrix(home, away, model, homeAdv);
const { lambdaHome, lambdaAway } = matrixLambdas(m);
return {
home,
away,
eloHome,
eloAway,
eloHome: ratingOf(model, home),
eloAway: ratingOf(model, away),
lambdaHome,
lambdaAway,
probs: outcomeProbs(m),
+91
View File
@@ -0,0 +1,91 @@
import { describe, it, expect } from 'vitest';
import { fitTeamDc, teamDcLambdas, poolMatrices, type DcMatch } from './teamDc';
import { scoreMatrix } from './poisson';
function synth(): DcMatch[] {
// Strong beats Weak repeatedly; Mid splits with both.
const ms: DcMatch[] = [];
for (let i = 0; i < 30; i++) {
ms.push({ daysAgo: i * 10, home: 'Strong', away: 'Weak', homeGoals: 3, awayGoals: 0, neutral: true });
ms.push({ daysAgo: i * 10, home: 'Mid', away: 'Strong', homeGoals: 0, awayGoals: 2, neutral: true });
ms.push({ daysAgo: i * 10, home: 'Weak', away: 'Mid', homeGoals: 0, awayGoals: 2, neutral: true });
}
return ms;
}
describe('fitTeamDc', () => {
const p = fitTeamDc(synth(), 1, 5);
it('ranks attack and defence sensibly', () => {
expect(p.attack['Strong']!).toBeGreaterThan(p.attack['Mid']!);
expect(p.attack['Mid']!).toBeGreaterThan(p.attack['Weak']!);
expect(p.defence['Strong']!).toBeLessThan(p.defence['Weak']!); // lower = concedes less
});
it('keeps strengths normalized around 1', () => {
const atts = Object.values(p.attack);
const mean = atts.reduce((s, x) => s + x, 0) / atts.length;
expect(mean).toBeCloseTo(1, 6);
expect(p.mu).toBeGreaterThan(0.5);
expect(p.mu).toBeLessThan(3);
});
it('predicts higher λ for the stronger side and falls back for unknowns', () => {
const { lambdaHome, lambdaAway } = teamDcLambdas(p, 'Strong', 'Weak', true);
expect(lambdaHome).toBeGreaterThan(lambdaAway);
const fb = teamDcLambdas(p, 'Nowhere FC', 'Strong', true);
expect(fb.lambdaHome).toBeGreaterThan(0); // unseen team gets average strengths
});
it('shrinkage pulls thin histories toward average', () => {
const thin: DcMatch[] = [{ daysAgo: 1, home: 'A', away: 'B', homeGoals: 5, awayGoals: 0, neutral: true }];
const loose = fitTeamDc(thin, 0, 1);
const tight = fitTeamDc(thin, 0, 50);
expect(Math.abs(tight.attack['A']! - 1)).toBeLessThan(Math.abs(loose.attack['A']! - 1));
});
it('home multiplier learned from non-neutral games', () => {
const ms: DcMatch[] = [];
for (let i = 0; i < 60; i++) {
ms.push({ daysAgo: i, home: 'X', away: 'Y', homeGoals: 2, awayGoals: 1, neutral: false });
ms.push({ daysAgo: i, home: 'Y', away: 'X', homeGoals: 2, awayGoals: 1, neutral: false });
}
const hp = fitTeamDc(ms, 0, 5);
expect(hp.homeMult).toBeGreaterThan(1.05); // symmetric fixtures, home side scores more
});
});
describe('ensemble predictMatch', () => {
it('pools Elo-DC and Team-DC and stays a valid distribution', async () => {
const { predictMatch } = await import('./predict');
const p = fitTeamDc(synth(), 1, 5);
const model = {
asOf: '2026-06-10',
params: { goalsPerElo: 0.0057, avgGoals: 2.73, rho: -0.05, homeAdvElo: 70 },
ratings: { Strong: 2000, Weak: 1700 },
teamDc: p,
ensembleW: 0.75,
};
const pred = predictMatch('Strong', 'Weak', model);
expect(pred.probs.home + pred.probs.draw + pred.probs.away).toBeCloseTo(1, 6);
expect(pred.probs.home).toBeGreaterThan(0.5);
expect(pred.lambdaHome).toBeGreaterThan(pred.lambdaAway);
// pure-Elo path still works when ensemble params are absent
const pure = predictMatch('Strong', 'Weak', { ...model, teamDc: undefined as never, ensembleW: undefined as never });
expect(pure.probs.home + pure.probs.draw + pure.probs.away).toBeCloseTo(1, 6);
});
});
describe('poolMatrices', () => {
it('renormalizes and interpolates between the two models', () => {
const a = scoreMatrix(2.0, 0.8, -0.05);
const b = scoreMatrix(1.0, 1.4, -0.05);
const pooled = poolMatrices(a, b, 0.5);
let sum = 0;
for (const row of pooled) for (const v of row) sum += v;
expect(sum).toBeCloseTo(1, 9);
// w=1 returns A, w=0 returns B (up to renormalization)
const pa = poolMatrices(a, b, 1);
expect(pa[2]![0]!).toBeCloseTo(a[2]![0]!, 9);
});
});
+141
View File
@@ -0,0 +1,141 @@
// Team-level Dixon-Coles: every nation gets multiplicative attack/defence
// strengths, fit by weighted MLE (Maher-style iterative updates) with
// exponential time-decay and shrinkage toward the mean for thin histories.
// This is the second, Elo-independent member of the ensemble.
export interface TeamDcParams {
/** Mean goals per team per match (league baseline). */
mu: number;
/** Home-ground multiplier applied to the home side's λ (1 on neutral). */
homeMult: number;
attack: Record<string, number>;
defence: Record<string, number>;
}
export interface DcMatch {
/** Days before the fit date (for decay weighting). */
daysAgo: number;
home: string;
away: string;
homeGoals: number;
awayGoals: number;
neutral: boolean;
}
/**
* Fit attack/defence strengths.
* @param xi decay per year (weight = exp(-xi * daysAgo/365)); 0 = no decay
* @param shrinkK pseudo-observations pulling strengths toward 1.0
*/
export function fitTeamDc(matches: DcMatch[], xi: number, shrinkK: number, iterations = 25): TeamDcParams {
const w = matches.map((m) => Math.exp((-xi * m.daysAgo) / 365));
const teams = new Set<string>();
for (const m of matches) { teams.add(m.home); teams.add(m.away); }
// initial values
let mu = 0;
{
let g = 0, n = 0;
for (let i = 0; i < matches.length; i++) {
g += w[i]! * (matches[i]!.homeGoals + matches[i]!.awayGoals);
n += w[i]! * 2;
}
mu = n > 0 ? g / n : 1.3;
}
let homeMult = 1.15;
const att = new Map<string, number>();
const def = new Map<string, number>();
for (const t of teams) { att.set(t, 1); def.set(t, 1); }
for (let it = 0; it < iterations; it++) {
// attack updates
const num = new Map<string, number>();
const den = new Map<string, number>();
for (let i = 0; i < matches.length; i++) {
const m = matches[i]!;
const wi = w[i]!;
const hF = m.neutral ? 1 : homeMult;
num.set(m.home, (num.get(m.home) ?? 0) + wi * m.homeGoals);
den.set(m.home, (den.get(m.home) ?? 0) + wi * mu * def.get(m.away)! * hF);
num.set(m.away, (num.get(m.away) ?? 0) + wi * m.awayGoals);
den.set(m.away, (den.get(m.away) ?? 0) + wi * mu * def.get(m.home)!);
}
for (const t of teams) {
att.set(t, ((num.get(t) ?? 0) + shrinkK * mu) / ((den.get(t) ?? 0) + shrinkK * mu));
}
// defence updates
num.clear(); den.clear();
for (let i = 0; i < matches.length; i++) {
const m = matches[i]!;
const wi = w[i]!;
const hF = m.neutral ? 1 : homeMult;
// home defence faces away attack
num.set(m.home, (num.get(m.home) ?? 0) + wi * m.awayGoals);
den.set(m.home, (den.get(m.home) ?? 0) + wi * mu * att.get(m.away)!);
// away defence faces home attack (with home boost)
num.set(m.away, (num.get(m.away) ?? 0) + wi * m.homeGoals);
den.set(m.away, (den.get(m.away) ?? 0) + wi * mu * att.get(m.home)! * hF);
}
for (const t of teams) {
def.set(t, ((num.get(t) ?? 0) + shrinkK * mu) / ((den.get(t) ?? 0) + shrinkK * mu));
}
// normalize so mean(att) = mean(def) = 1 (identifiability)
const aMean = [...att.values()].reduce((s, x) => s + x, 0) / att.size;
const dMean = [...def.values()].reduce((s, x) => s + x, 0) / def.size;
for (const t of teams) { att.set(t, att.get(t)! / aMean); def.set(t, def.get(t)! / dMean); }
mu = mu * aMean * dMean;
// home multiplier from non-neutral games
let hNum = 0, hDen = 0;
for (let i = 0; i < matches.length; i++) {
const m = matches[i]!;
if (m.neutral) continue;
hNum += w[i]! * m.homeGoals;
hDen += w[i]! * mu * att.get(m.home)! * def.get(m.away)!;
}
if (hDen > 0) homeMult = Math.min(1.6, Math.max(1, hNum / hDen));
}
return {
mu,
homeMult,
attack: Object.fromEntries(att),
defence: Object.fromEntries(def),
};
}
/** λs for a fixture; unseen teams fall back to average (1.0) strengths. */
export function teamDcLambdas(
p: TeamDcParams,
home: string,
away: string,
neutral: boolean,
): { lambdaHome: number; lambdaAway: number } {
const ah = p.attack[home] ?? 1;
const dh = p.defence[home] ?? 1;
const aa = p.attack[away] ?? 1;
const da = p.defence[away] ?? 1;
const hF = neutral ? 1 : p.homeMult;
return {
lambdaHome: Math.min(8, Math.max(0.15, p.mu * ah * da * hF)),
lambdaAway: Math.min(8, Math.max(0.15, p.mu * aa * dh)),
};
}
/** Cell-wise log-opinion pool of two score matrices: mA^w · mB^(1-w), renormalized. */
export function poolMatrices(mA: number[][], mB: number[][], w: number): number[][] {
const out: number[][] = [];
let total = 0;
for (let i = 0; i < mA.length; i++) {
out[i] = [];
for (let j = 0; j < mA[i]!.length; j++) {
const v = Math.pow(Math.max(1e-12, mA[i]![j]!), w) * Math.pow(Math.max(1e-12, mB[i]![j]!), 1 - w);
out[i]![j] = v;
total += v;
}
}
for (const row of out) for (let j = 0; j < row.length; j++) row[j]! /= total;
return out;
}