方法论 · Methodology

rolling-20 goal window:
  attack_strength  = goals_for_20     / (20 × league_avg_home_goals)
  defense_strength = goals_against_20 / (20 × league_avg_away_goals)

Bayesian shrinkage toward league mean 1.0:
  adjusted = (k × 1.0 + n × raw) / (k + n)
  k = 5   (shrinkage parameter)

ELO blend:
  strength_final = w × adjusted_rolling + (1 − w) × f(elo_diff)

λ_home = home_attack_strength × away_defense_strength × league_avg_home_goals
λ_away = away_attack_strength × home_defense_strength × league_avg_away_goals

league_avg_* 来自 load_league_avgs() 实时查询（非 fixture 列）

P(home = i, away = j) = Poisson(i ; λ_home) × Poisson(j ; λ_away)

SCORE_MAX = 7      (i, j ∈ {0, 1, ..., 7})
After grid: renormalize so sum(P) = 1

P(home)  = Σ_{i > j}  P(i, j)
P(draw)  = Σ_{i = j}  P(i, j)
P(away)  = Σ_{i < j}  P(i, j)

AH lines: 0.00 · -0.25 · -0.50 · -0.75 · -1.00

win_credit_prob = P(FULL_WIN) + 0.5 × P(HALF_WIN)

OU k.5:
  P(over)  = Σ_{i + j > k} P(i, j)
  P(under) = 1 − P(over)           ← 规范实现

BTTS:
  P(yes) = 1 − P(home = 0) − P(away = 0) + P(0, 0)
  P(no)  = 1 − P(yes)

p_calibrated = sigmoid(a × p_raw + b)
             = 1 / (1 + exp(−(a × p_raw + b)))

optimizer:
  Newton 法 · 2 × 2 Hessian · max_iter = 200
  train on rolling n_train settled matches

edge = model_prob − market_fair_prob

support_count = |{ providers | provider ≠ 'model_poisson'
                   AND fair_prob 可用 AND 数据新鲜 }|

accept(signal) ⇔
    support_count ≥ threshold_support
  AND |edge|        ≥ threshold_edge
  AND freshness     ≤ max_latency
  AND trigger_condition(signal)  is TRUE
  AND provider      ≠ 'model_poisson'

consensus_timeseries: AVG(fair_prob) per snapshot_ts

opening: 最早时间戳（> 24h before kickoff）
mid:     最接近 kickoff 前 6h 的快照
pre:     kickoff 前最近快照

delta_open_mid = p_mid − p_open
delta_mid_pre  = p_pre − p_mid
delta_open_pre = delta_open_mid + delta_mid_pre

abs_move       = | delta_open_pre |
steam_k        = 加权最近 3 个 delta_p   (k = 3)
ols_slope      = OLS 斜率(p_t ~ time_to_kickoff)
late_move_sign = sign(delta_mid_pre)
reversal_flag  = 1 if sign(d_om) ≠ sign(d_mp) AND 幅度显著
volatility     = std(delta_p)

lineup_quality_delta : 首发 XI 相对 rolling-20 的质量偏差

λ sensitivity:
  β_att = 0.06
  β_def = 0.03
  (HOLD_FROZEN · n = 23 / 60)

P3-A7 signal subset:
  S_signal = { matches | z_total ≠ 0 }
  S_zero   = { matches | z_total = 0 }

xGOT_home = Σ player_xGOT_home × λ_weight(α, β)
xGOT_away = Σ player_xGOT_away × λ_weight(α, β)

λ_nextgen = λ_base × (1 + α × (xGOT / league_avg_xGOT − 1))

P_xgot(i, j) = Poisson(i ; λ_xgot_home) × Poisson(j ; λ_xgot_away)

Best params: α = 0.05 · β = 0.05
Brier gap:   −0.000700  vs 独立 Poisson

signal_candidate_registry:
  module_id   · variant   · status
  verdict     · reviewer  · decided_at

forward_gate_registry_snapshot:
  module_key  · status_bucket
  settled_n   · required_n   · ready_flag
  last_eval   · action

Brier Score = (1/n) × Σ (p − y)²

LogLoss     = −(1/n) × Σ [ y × log(p) + (1 − y) × log(1 − p) ]

ECE (10-bin) = Σ_k (n_k / n) × | avg_p_k − acc_k |

X : 预测概率（分 10 桶）
Y : 实际命中率（该桶内样本的真阳占比）
对角线 y = x  ≡  完美校准
偏离方向 → 过信 / 欠信的诊断