live · review room
Every edge goes to review.
Three football models that check themselves against the bookmakers.
Three independent models, one per market, each trained separately for every league.
VAR puts probabilities on football matches in three markets: match outcome, both teams to score, and total goals. Each market has its own model, and each league its own champion per model.
The bookmakers' closing odds never enter the features. With the margin stripped out, they become the benchmark: every model is scored against the de-vigged closing line on the same matches.
There is no neural network anywhere, only gradient-boosted trees and Poisson rating models.
It exists to find out whether the models know anything the market doesn't, so most of the work goes into the checking.
The name: variance, and the video assistant referee.
Pick a camera, pick a league. The diagram redraws from that league's champion, and one example fixture runs through all three.
Champions board · 14 leagues · Tempo
Tempo: features selected, then the fitted Poisson pool weight w. Left is all XGBoost, right is all Poisson.
championLa Liga · N 1,900 · ensemble · F 32 · w 0.53
Branch A: 12 XGBoost members
Features
X: (1,900, 32)
raw (N_raw, 101) · home-minus-away deltas, shift(1)-lagged per team
4 regimes × 3 seeds
(B, F) → (B, 3)
multi:softprob, num_class=3
slow deep: lr 0.018 · 764 rounds · depth 6 · gamma base
width = rounds, height = depth, one outline per seed (42, 1042, 2042)
Uniform mean
(B, 3)
stack (12, B, 3)
2,257 rounds × 3 seeds × 3 classes = 20,313 trees
Branch B: Poisson member
Team ids
(B, 2)
Rates λ
(B, 2)
λ_home = exp(μ + home_adv + att[h] − def[a])
λ_away = exp(μ + att[a] − def[h])
L-BFGS-B, time decay ξ = 0.002/day, no L2
2 + 2T + 1 scalars · T = 26
example fixture · drag
Score grid
(B, 13, 13)
1 : 1 · 13.0% · likeliest score
0 to 12 goals. Brightness is probability.
Outlined: the 4 low-score cells with the Dixon-Coles correction. Then renormalise.
Dixon-Coles ρ fitted second-stage, bounded to [−0.3, 0.3]
Read-out
(B, 3)
triangle sums, for the example fixture:
Log-opinion pool
(B, 3)
p ∝ p_xgb^(1−w) · p_poisson^w
log-opinion pool, row-normalised. w is one scalar fitted on out-of-fold rows.
w = 0.53
Calibration
(B, 3) = P(home), P(draw), P(away)
softmax(log p / T)
temperature scaling, one scalar, bounds [0.25, 4.0]
Drawn: what T does to one probability, at four temperatures.
A challenger has to win the replay, then clear every check.
Walk-forward CV · expanding by season · 3 folds
Minimum two training seasons. A partial season is dropped before any of this starts.
Registry · step 1 · rank
10,000 resamples
≥ 95% to pass
one dot per resample. the challenger needs the yellow ones.
Registry · step 2 · gates
check completeChecking: Tempo
Checking: Blaze
Checking: Storm
vs climatology
vs the market
not asked
calibration
returns
not asked
fold stability
own market
Checking: Tempo
Checking: Blaze
Checking: Storm
A league's first run is seeded as champion regardless of gates.
Three lanes. A bar is something the models have in common. A cut is something that never crosses between them.
TTempo1X2
BBlazeBTTS
SStormO/U
One raw CSV per league feeds all three models.
One partial-season rule and one walk-forward CV scheme. Duplicated code, same logic.
The same Poisson rate formula in all three.
The same four regimes and three seeds in Tempo and Blaze.
The same registry rule.
No fitted object crosses models. Each league fits three separate Poisson models.
Tempo's and Blaze's Poisson fits use identical inputs and settings, so their ratings are equal. Each still computes its own.
Feature caches are separate. Blaze recomputes its three Tempo-style quality features itself.
No cross-package import. All three carry their own copy of the training logic.
(B, 3) · (B,) · (B, 5)
Slate / dashboard
The three outputs join in one place only: the slate and dashboard layer, per fixture.
Why it is built this way.
Folds go by season, in order, and are never shuffled. Every headline metric is pooled out-of-fold predictions.
Closing odds are deliberately kept out of the features. De-vigged, they are what every model is scored against, on the same matches.
Candidates are ranked on log loss alone, and every other metric is a pass-or-fail gate. An earlier composite score counted log loss and Brier as if they measured two different things.
Rules written down in advance are kept apart from rules the grid search found.
A champion lasts until a challenger beats it. The checks stay the same.
Results, experiments and failures live in Briefs and Logs.