Predictorous

What I rejected

Every idea below was built, tested against the shipping model on fixtures it had not seen, and lost. They are here because a model you can only see the successes of is a model you have no reason to trust — and because several of these are the ideas anyone would try next.

The test is a paired bootstrap on the fixtures an idea actually touches. P(better) is the share of resamples in which the new version beat the old one — about 0.5 is a coin flip. Nothing here got near enough to 0.95 to ship on its merits. One thing shipped anyway, on a single season, and the longer test took it back out; that entry is marked reversed, with the dates.

A team-quality layer inside the predictor

rejected+0.0001 log loss

The hope. Twenty-eight opponent-adjusted team metrics ought to know something about Saturday that goal-based ratings do not.

What happened. It moved the log loss by +0.0001. Not a small win — no win. The information was already in the chance rates the model reads.

The manager as a separate effect

rejected−0.0132 log loss when fully spell-scoped

The hope. A new manager changes a side, so a manager term should carry signal the club's own history does not.

What happened. Tested four independent ways and rejected each time. What survives is SELECTION — who he picks — and that is already visible in the squad. Scoping the whole history to the current spell made it actively worse.

Availability corrections to a club's history

not provenP(better) = 0.633

The hope. A side missing three regulars is not the side its record describes, so the history should be discounted.

What happened. The right shape — it has the property that a full-strength side gets no correction at all — but the effect could not be separated from noise on 232 fixtures.

The layer: let the Elo arm's classifier own the draw

reversedlayer 0.9850 v blend 0.9843, 30.1% v 31.6% of draws found

The hope. On the 1,714 fixtures of 25/26 the Elo arm ranked draws best, so average the two models for the home:away ratio and hand P(draw) to the Elo arm alone. It shipped on 2026-09-20.

What happened. The eight-season walk-forward (11,311 fixtures, every fold fitted only on earlier seasons) reversed it: the layer imports the classifier's draw ORDER, which is worse than the plain blend's, and loses on the triplet too — 0.9850 against 0.9843. The plain blend ships from 2026-11-09; rows priced by the layer stand on the record as recorded.

Blending the process-Elo arm into the shipped model

not proven0.9843 v 0.9842 log loss, 11,311 fixtures

The hope. Two independent models usually beat one.

What happened. On a few hundred fixtures it was better and nowhere near significant (P(better) 0.790). On the eight-season walk-forward the blend and the shipped model are a coin flip on log loss — 0.9843 against 0.9842 — and the blend finds more draws at the cutoff the product uses. It ships for that reason, not because it won.

Seeding the model with a manager's record

not provenP(better) = 0.840

The hope. Same idea as above, from the other end.

What happened. Closer, still short. Same answer: wait for the data.

Context features for the draw model

rejected28.9% vs 29.6%, inside noise

The hope. Points-per-game gap, how far into the season it is, whether both sides are comfortable — a drawish match ought to be predictable from its situation.

What happened. The model assigned them coefficients of ±0.001 to 0.04 and out-of-sample performance did not move. Consistent with the separate finding that drawish weeks are random clumping: overdispersion measured 1.0 across ten league-seasons.

Zones built from individual players

rejected

The hope. If a side's zone profile is the sum of who plays in it, building it from players should beat building it from team totals.

What happened. It did not beat the shipping model. The team totals already had it.

“Volume beats weighting” — my own claim

reversed3.77 projected vs 2.75 actual

The hope. I argued that how much a side creates matters more than how the chances are weighted, and had a result to show for it.

What happened. The result came from a conversion constant I had fitted by log loss instead of solving. Projections came out at 3.77 goals a game against an actual 2.75. Solving it properly INVERTED the ranking and falsified the claim.

What this leaves

A shipping model that, over eight frozen seasons and 11,311 fixtures, scores 0.984 log loss against a base rate of 1.073, calls the right outcome 52.7% of the time where always picking the home side gets 43.4%, and is calibrated where the mass is. Everything above was an attempt to improve on that and did not.

The honest constraint is sample size. Most of these were tested on a few hundred fixtures, which cannot separate ideas that differ by a few thousandths of log loss. They are not dead — they are unproven, and the difference matters. Four more leagues are being collected for exactly this reason.

The track record is what the model that survived all this has actually done.