🧪 模型实验室
回测 · 训练 · 监控当前视图:监控信号。长任务完成后将在此处显示结果。
📈 回测中心
自动复盘流程:加载快照 → 自动获取赛果 → 逐场预测比对 → 规则有效性评估 → 参数调优建议
复盘结果保存至
复盘结果保存至
reports/ 目录,并持久化规则性能数据供趋势追踪。🔍 执行回测
📊 赛果回填
手动触发赛果自动获取(含天天盈球、500彩票网、足彩狗、澳客、预测家与 openfootball):
🤖 模型训练中心
📊 训练数据概览
历史训练数据来源:MatchPredict(2023-24 赛季)+ football-data.co.uk(2024-25 / 2025-26 两赛季五大联赛,含亚盘盘口与水位特征)+ SQLite 已有赛果;特征维度 18 维。
暂无训练数据
⚙️ 训练配置
📡 监控信号
observation-only · 只读不下注
ℹ️ 说明
本面板集中展示未经样本外证据升级的观测型信号(串关闭环、凯利离散度、Dixon-Coles 比分回测、周度 binomial 一致性、市场边际地图)。这些信号仅供研究观测,不改变任何仓位或下注决策,与详情页决策核心物理隔离。
🎫 串关闭环追踪
只读 parlay_records1428
累计串关(目标 60)
72
覆盖天数(目标 30)
1279/149
已结算 / 待结算
7.1%
命中率(已结算)
样本量已达统计门槛,命中率/ROI 可作参考(仍为小仓位娱乐口径)。
| market/玩法 | 条数 | 已结算 | 胜/负 | 命中率 | 盈亏 | ROI |
|---|---|---|---|---|---|---|
| beidan/double | 1 | 1 | 0/1 | 0.0% | -30.00 | -100.0% |
| beidan/goals | 69 | 61 | 0/61 | 0.0% | -488.00 | -100.0% |
| beidan/hcap | 248 | 231 | 15/216 | 6.5% | +10347.76 | +559.9% |
| beidan/htft | 1 | 0 | 0/0 | — | +0.00 | — |
| beidan/score | 69 | 63 | 1/62 | 1.6% | -205.44 | -40.8% |
| beidan/single | 1 | 1 | 0/1 | 0.0% | -20.00 | -100.0% |
| beidan/spf | 373 | 277 | 30/247 | 10.8% | +1764.80 | +75.3% |
| jingcai/double | 2 | 2 | 0/2 | 0.0% | -60.00 | -100.0% |
| jingcai/goals | 65 | 65 | 0/61 | 0.0% | -488.00 | -93.8% |
| jingcai/hcap | 233 | 225 | 19/203 | 8.4% | -662.48 | -36.8% |
| jingcai/htft | 3 | 3 | 1/2 | 33.3% | +57.00 | +38.0% |
| jingcai/score | 64 | 64 | 0/58 | 0.0% | -464.00 | -90.6% |
| jingcai/single | 2 | 2 | 2/0 | 100.0% | +12.60 | +31.5% |
| jingcai/spf | 297 | 284 | 23/258 | 8.1% | -472.50 | -19.0% |
半全场为模型折算口径(非官方兑付),与胜平负官方盘口分列。
📐 周度一致性监控(binomial)
实盘 vs 回测预期窗口 7 天 · market=jingcai · 样本 326 行 · 状态: 一致
| cohort | 判定 |
|---|---|
| all | ✅ [all] 一致:实盘 0.5566 (95%CI 0.4893~0.6219) 覆盖回测预期 0.5223,n=212 |
| standard | ✅ [standard] 样本不足 n=0<30,暂不判定(实盘 NA vs 预期 0.7221) |
跳过:无方向 0 / 放弃 0 / 赛果未齐 114
🗺️ 市场边际地图
市场边际地图-20260829.json全量样本外分层命中率/ROI,用于观测「哪些切片系统性有边际」。全集期望值为负属正常(庄家抽水),高 ROI 冷门切片样本极小不可外推。
47.7%
胜平负 命中率(n=24731, ROI -6.0%)
51.6%
让球 命中率(n=909, ROI -10.7%)
胜平负 · 按赔率段
| 切片 | n | 命中率 | ROI |
|---|---|---|---|
| >=5.0(冷门) | 4 | 25.0% | +81.9% |
| 3.0-5.0 | 103 | 35.0% | +23.6% |
| <2.0(大热) | 8245 | 64.0% | -4.5% |
| 2.0-3.0 | 16379 | 39.6% | -7.1% |
胜平负 · 按置信段
| 切片 | n | 命中率 | ROI |
|---|---|---|---|
| 0.40-0.50 | 2488 | 46.3% | -3.8% |
| >=0.60 | 4741 | 70.2% | -4.2% |
| 0.50-0.60 | 3256 | 56.1% | -4.6% |
| <0.40 | 14246 | 38.6% | -7.8% |
胜平负 · 按方向
| 切片 | n | 命中率 | ROI |
|---|---|---|---|
| 负 | 8155 | 46.7% | -5.4% |
| 平 | 2522 | 34.3% | -5.8% |
| 胜 | 14054 | 50.7% | -6.4% |
胜平负 · 按联赛层级
| 切片 | n | 命中率 | ROI |
|---|---|---|---|
| 五大联赛 | 3439 | 48.0% | -4.0% |
| 其他 | 21292 | 47.7% | -6.3% |
胜平负 · 赔率段×联赛层级
| 切片 | n | 命中率 | ROI |
|---|---|---|---|
| 3.0-5.0|五大 | 4 | 50.0% | +82.8% |
| >=5.0(冷门)|其他 | 4 | 25.0% | +81.9% |
| 3.0-5.0|其他 | 99 | 34.3% | +20.4% |
| 2.0-3.0|五大 | 2307 | 40.4% | -3.8% |
| <2.0(大热)|其他 | 7117 | 64.1% | -4.5% |
| <2.0(大热)|五大 | 1128 | 63.4% | -4.7% |
| 2.0-3.0|其他 | 14072 | 39.5% | -7.7% |
实盘追踪:已落地 55 场(已结算 55),总命中率 54.5%。样本极小,仅供对照观测。
🎲 凯利离散度(KD)
kelly_dispersion_synthetic_20260827.md · 生产默认关Stage2 Kelly 离散度 focused synthetic report
This report is synthetic only. It is not a football-data historical backtest and makes no performance claim.
Safety contract
- Input is structured bookmaker 1X2 odds only.
Avg/Max-style aggregates and malformed odds are excluded.- Only matching
handicap=0and one timestamp are eligible; the synthetic timestamp stands in for T−24 and is not a source publication timestamp. - At least three valid independent bookmakers are required.
*_stdis population standard deviation (denominator n); both KD thresholds usekelly_max_stdacross directions (no average-std fallback).- KD-LOW emits a dynamic consensus variant only; KD-HIGH emits a risk label with no outcome.
- Both KD rules are observation-only and forced to D; even a high synthetic statistic is never reported as recommended enablement.
Synthetic observations
| fixture | valid companies | eligible | prob overall std | Kelly max std | overall range | variants |
|---|---|---|---|---|---|---|
| --- | ---: | :---: | ---: | ---: | ---: | --- |
| low-consensus | 3 | True | 0.0 | 0.0 | 0.0 | KD-LOW-v1=win |
| high-dispersion | 3 | True | 0.133065 | 0.940793 | 0.455717 | KD-HIGH-v1=high-dispersion |
| aggregate-and-malformed | 3 | True | 0.0 | 0.0 | 0.0 | KD-LOW-v1=win |
Interpretation
- This fixture confirms the safety gates and variant plumbing only.
- No historical hit rate, EV, evidence grade upgrade, or production recommendation is inferred.
config.yamlkeepsrules.kelly_dispersion.enabled: false; the current main decision remains unchanged.- The production registry currently receives no company_odds from the main ensemble path; KD requires an explicit structured panel and therefore remains observation-only.
📊 Dixon-Coles τ 比分回测(DC)
DC比分回测-20260828.md · 生产默认关Dixon-Coles τ 回测报告(M2)
日期:2026-08-28 | ρ=-0.08 (config
dc.rho) | ξ 时间衰减未接入 (M3)一、概览
- CSV: 112 个; 有效: 32641; 剔除(<3家): 6726
- 面板分布: 3×4364, 4×5201, 5×4771, 6×18305
二、比分 top-5
| 方法 | 场次 | 命中 | 命中率 | Δpp |
|---|---|---|---|---|
| ------ | ------ | ------ | -------- | ----- |
| 独立泊松 | 32641 | 16588 | 50.8% | — |
| DC | 32641 | 16697 | 51.2% | ++0.33 |
- 目标 +≥1pp → ❌ 未达标
三、大小球 2.5
| 方法 | 命中 | 命中率 | Δpp |
|---|---|---|---|
| ------ | ------ | -------- | ----- |
| 独立泊松 | 17943 | 55.0% | — |
| DC | 18413 | 56.4% | ++1.44 |
四、裁决
- 保持 dc.enabled=false (比分提升 +0.33pp)