model_extended.js

/**
 * 今彩539 進階模型搜尋
 * 測試 15+ 種不同策略，目標是找到最高「至少命中 1 個」的模型
 *
 * 理論上限 (5 picks from 39):
 *   P(≥1 hit) = 1 - C(34,5)/C(39,5) ≈ 51.64%
 * 理論上限 (7 picks from 39):
 *   P(≥1 hit) = 1 - C(32,5)/C(39,5) ≈ 64.29%
 * 理論上限 (8 picks from 39):
 *   P(≥1 hit) = 1 - C(31,5)/C(39,5) ≈ 69.42%
 */

const fs = require('fs');

const raw = JSON.parse(fs.readFileSync('lottery_539_data.json', 'utf8'));
const draws = raw.draws;
const TOTAL = draws.length;
const BACKTEST_N = 1000;
const START_IDX = TOTAL - BACKTEST_N;

function pad(n) { return String(parseInt(n)).padStart(2, '0'); }

// ─── Shared Helpers ───────────────────────────────────────────────────────────
function getFreq(draws, upTo) {
  const f = {};
  for (let i = 1; i <= 39; i++) f[pad(i)] = 0;
  for (let i = 0; i < upTo; i++)
    for (const n of draws[i].numbers) f[pad(n)]++;
  return f;
}

function getGaps(draws, upTo, window = 600) {
  const g = {};
  for (let i = 1; i <= 39; i++) g[pad(i)] = upTo;
  for (let i = upTo - 1; i >= Math.max(0, upTo - window); i--)
    for (const n of draws[i].numbers)
      if (g[pad(n)] === upTo) g[pad(n)] = upTo - 1 - i;
  return g;
}

function getRecentFreq(draws, upTo, win = 30) {
  const f = {};
  for (let i = 1; i <= 39; i++) f[pad(i)] = 0;
  for (let i = Math.max(0, upTo - win); i < upTo; i++)
    for (const n of draws[i].numbers) f[pad(n)]++;
  return f;
}

// Exponential-decay frequency (recent draws weighted more)
function getDecayFreq(draws, upTo, halfLife = 100) {
  const f = {};
  for (let i = 1; i <= 39; i++) f[pad(i)] = 0;
  for (let i = Math.max(0, upTo - 800); i < upTo; i++) {
    const age = upTo - 1 - i;
    const weight = Math.exp(-Math.LN2 * age / halfLife);
    for (const n of draws[i].numbers) f[pad(n)] += weight;
  }
  return f;
}

// Number pair co-occurrence: how often each number appears with the most recent draw's numbers
function getPairScore(draws, upTo, window = 300) {
  const lastNums = upTo > 0 ? new Set(draws[upTo - 1].numbers.map(n => pad(n))) : new Set();
  const pair = {};
  for (let i = 1; i <= 39; i++) pair[pad(i)] = 0;
  for (let i = Math.max(0, upTo - window - 1); i < upTo - 1; i++) {
    const nums = draws[i].numbers.map(n => pad(n));
    const hasCommon = nums.some(n => lastNums.has(n));
    if (hasCommon) {
      const nextIdx = i + 1;
      if (nextIdx < upTo) {
        for (const n of draws[nextIdx].numbers) pair[pad(n)]++;
      }
    }
  }
  return pair;
}

// Numbers NOT seen in last k draws
function getAvoidRecent(draws, upTo, k = 5) {
  const recent = new Set();
  for (let i = Math.max(0, upTo - k); i < upTo; i++)
    for (const n of draws[i].numbers) recent.add(pad(n));
  return recent;
}

const BUCKETS_5 = [[1,2,3,4,5,6,7,8],[9,10,11,12,13,14,15],[16,17,18,19,20,21,22],[23,24,25,26,27,28,29,30],[31,32,33,34,35,36,37,38,39]];
const BUCKETS_7 = [[1,2,3,4,5,6],[7,8,9,10,11],[12,13,14,15,16],[17,18,19,20,21],[22,23,24,25,26,27],[28,29,30,31,32],[33,34,35,36,37,38,39]];
const BUCKETS_8 = [[1,2,3,4,5],[6,7,8,9,10],[11,12,13,14,15],[16,17,18,19,20],[21,22,23,24,25],[26,27,28,29,30],[31,32,33,34,35],[36,37,38,39]];

function topByScore(scores, n) {
  return Object.entries(scores).sort((a,b)=>b[1]-a[1]).slice(0,n).map(([k])=>k);
}

// ─── Models ───────────────────────────────────────────────────────────────────
const MODELS = {
  // Baseline: Current best
  M3_SpreadOverdue_5: (draws, i) => {
    const g = getGaps(draws, i);
    return BUCKETS_5.map(b => b.map(n=>({n:pad(n),s:g[pad(n)]||0})).sort((a,b)=>b.s-a.s)[0].n);
  },

  // 7 numbers: most-overdue from 7 equal buckets
  M_Spread7: (draws, i) => {
    const g = getGaps(draws, i);
    return BUCKETS_7.map(b => b.map(n=>({n:pad(n),s:g[pad(n)]||0})).sort((a,b)=>b.s-a.s)[0].n);
  },

  // 8 numbers: most-overdue from 8 buckets
  M_Spread8: (draws, i) => {
    const g = getGaps(draws, i);
    return BUCKETS_8.map(b => b.map(n=>({n:pad(n),s:g[pad(n)]||0})).sort((a,b)=>b.s-a.s)[0].n);
  },

  // Top 5 by exponential decay frequency (recent = high score)
  M_DecayFreq5: (draws, i) => {
    const df = getDecayFreq(draws, i, 80);
    // Invert: pick LEAST seen recently (cold by decay)
    const inv = {};
    const max = Math.max(...Object.values(df)) || 1;
    for (const k in df) inv[k] = max - df[k];
    return topByScore(inv, 5);
  },

  // Composite: gap(50%) + cold-decay(30%) + avoids last-5-draws(20%)
  M_CompositeAvoid5: (draws, i) => {
    const g = getGaps(draws, i);
    const df = getDecayFreq(draws, i, 80);
    const avoid = getAvoidRecent(draws, i, 5);
    const maxG = Math.max(...Object.values(g)) || 1;
    const maxDf = Math.max(...Object.values(df)) || 1;
    const sc = {};
    for (let n = 1; n <= 39; n++) {
      const k = pad(n);
      sc[k] = (g[k]/maxG)*50 + ((maxDf-df[k])/maxDf)*30 + (avoid.has(k) ? 0 : 20);
    }
    return topByScore(sc, 5);
  },

  // Same as above but 7 picks
  M_CompositeAvoid7: (draws, i) => {
    const g = getGaps(draws, i);
    const df = getDecayFreq(draws, i, 80);
    const avoid = getAvoidRecent(draws, i, 5);
    const maxG = Math.max(...Object.values(g)) || 1;
    const maxDf = Math.max(...Object.values(df)) || 1;
    const sc = {};
    for (let n = 1; n <= 39; n++) {
      const k = pad(n);
      sc[k] = (g[k]/maxG)*50 + ((maxDf-df[k])/maxDf)*30 + (avoid.has(k) ? 0 : 20);
    }
    return topByScore(sc, 7);
  },

  // Spread 7 + pair-correlation bonus
  M_Spread7Pair: (draws, i) => {
    const g = getGaps(draws, i);
    const p = getPairScore(draws, i, 200);
    const maxP = Math.max(...Object.values(p)) || 1;
    const sc = {};
    for (let n = 1; n <= 39; n++) {
      const k = pad(n);
      sc[k] = (g[k] || 0) * 0.7 + (p[k]/maxP) * 100 * 0.3;
    }
    return BUCKETS_7.map(b => b.map(n=>({n:pad(n),s:sc[pad(n)]||0})).sort((a,b)=>b.s-a.s)[0].n);
  },

  // Short-window overdue (50 draws) — focus on very recent absence
  M_ShortGap5: (draws, i) => {
    const g = getGaps(draws, i, 50);
    return BUCKETS_5.map(b => b.map(n=>({n:pad(n),s:g[pad(n)]||0})).sort((a,b)=>b.s-a.s)[0].n);
  },

  // Short-window overdue: 7 picks
  M_ShortGap7: (draws, i) => {
    const g = getGaps(draws, i, 50);
    return BUCKETS_7.map(b => b.map(n=>({n:pad(n),s:g[pad(n)]||0})).sort((a,b)=>b.s-a.s)[0].n);
  },

  // Random baseline (for comparison)
  M_Random5: (draws, i) => {
    const nums = Array.from({length:39},(_,i)=>pad(i+1));
    for (let i = nums.length-1; i>0; i--) {
      const j = Math.floor(Math.random()*(i+1));
      [nums[i],nums[j]] = [nums[j],nums[i]];
    }
    return nums.slice(0,5);
  },
};

// ─── Run backtest for all models ─────────────────────────────────────────────
console.log(`📊 進階模型搜尋：${Object.keys(MODELS).length} 個模型 × ${BACKTEST_N} 期\n`);

const modelResults = {};
for (const name of Object.keys(MODELS)) {
  modelResults[name] = { hits: 0, atLeast1: 0, atLeast2: 0, atLeast3: 0, total: 0 };
}

const startTime = Date.now();
for (let i = START_IDX; i < TOTAL; i++) {
  const actual = new Set(draws[i].numbers.map(n => pad(n)));
  for (const [name, fn] of Object.entries(MODELS)) {
    const pred = fn(draws, i);
    let h = 0;
    for (const p of pred) if (actual.has(p)) h++;
    const r = modelResults[name];
    r.total++;
    r.hits += h;
    if (h >= 1) r.atLeast1++;
    if (h >= 2) r.atLeast2++;
    if (h >= 3) r.atLeast3++;
  }
  if ((i - START_IDX + 1) % 200 === 0) {
    const done = i - START_IDX + 1;
    process.stdout.write(`\r  進度: ${done}/${BACKTEST_N} (${((done/BACKTEST_N)*100).toFixed(0)}%) — ${((Date.now()-startTime)/1000).toFixed(1)}s`);
  }
}

console.log('\n\n');

// ─── Report ───────────────────────────────────────────────────────────────────
const rows = Object.entries(modelResults).map(([name, r]) => ({
  name,
  picks: name.includes('8') ? 8 : name.includes('7') ? 7 : 5,
  at1: ((r.atLeast1 / r.total) * 100).toFixed(2),
  at2: ((r.atLeast2 / r.total) * 100).toFixed(2),
  at3: ((r.atLeast3 / r.total) * 100).toFixed(2),
  hitRate: ((r.hits / (r.total * (name.includes('8') ? 8 : name.includes('7') ? 7 : 5))) * 100).toFixed(2),
})).sort((a, b) => parseFloat(b.at1) - parseFloat(a.at1));

console.log('模型名稱'.padEnd(30) + ' 預測數  ≥1命中%   ≥2命中%   ≥3命中%  整體命中%');
console.log('─'.repeat(80));
for (const r of rows) {
  const flag = r.picks > 5 ? ' ← 多選' : '';
  console.log(
    r.name.padEnd(30) +
    String(r.picks).padStart(5) +
    String(r.at1 + '%').padStart(10) +
    String(r.at2 + '%').padStart(10) +
    String(r.at3 + '%').padStart(10) +
    String(r.hitRate + '%').padStart(11) + flag
  );
}

console.log('\n理論上限: 5選 ≈51.64%  7選 ≈64.29%  8選 ≈69.42%');

fs.writeFileSync('model_extended.json', JSON.stringify(
  { rows, generatedAt: new Date().toISOString() }, null, 2), 'utf8');
console.log('\n✅ model_extended.json 已儲存');