★ ★ ★The★ ★ ★

AlmanacAnalysis

What it predicts, how, and how it went
—
Loading—

Clear numbers, plainly graded

Every football game the Almanac prices gets a chance for a home win, a draw and an away win. The biggest of the three is the top pick, and its chance is the confidence. This page puts the two side by side: what the model said before kick-off, and what happened. If calls quoted at 70% land about 70% of the time, the numbers can be trusted at face value. If they don't, this is where it shows.

Updated 15 November 2026. The tables below are live from the record; the prose describes the model as it stands on that date (the release log has every change since).

The banding table, 50% to 100%

Every graded football game with a top pick of 50% or more, grouped by its confidence in 5-point bands. Won counts the top pick landing, and nothing else. Read right also counts a game the model called as a draw scoreline (1-1, 0-0) that finished level. Each has its own quoted figure, so neither is flattered. Click a band to see its most recent games.

BandGamesWonWon %QuotedGapDrawsRead rightDaily List

Quoted against landed

The same bands as a picture. The dashed line is perfect calibration: a band on the line landed exactly as often as it was quoted. Above the line the model was too cautious, below it too bold. Bigger dots are bigger samples, so the small top bands deserve less weight than they appear to.

Day by day

Every graded day, newest first, for the games in one confidence bracket. It opens on 70% and above, the calls strong enough to matter. Pick another bracket, or set your own range. A bad day is one where the bracket landed well below its quoted chance, and it is usually a round full of draws and away wins.

DayGamesWonWon %QuotedDrawsRead right %

How a prediction is made

Football runs through one engine, engine.py, in seven stages. Nothing here is a black box: each number on a fixture row comes from a function you can read, and the settings below are read from the live build, not typed in.

Why it works this way

Built in on purpose

  • Ratings, not points tallies. Points already encode wins, goal difference uses an arbitrary scale, and neither is a probability. Attack and defence ratings feed a score model that produces probabilities directly.
  • Last season, then this season. Early in a season a few results say very little, so the prior carries the rating until enough games are in.
  • Calibration after the model. Raw Poisson output is over-confident at the bottom and cautious at the top. A two-number curve corrects it, refitted weekly and only shipped when it also improves a season it was not fitted on.
  • The model's own read only. No bookmaker prices go into any prediction. That keeps the list independent, at a known cost: see the next panel.
  • Cup ties on their own curve. Domestic cup favourites land far under a league favourite's number, so since October a second calibration pulls them down (a raw 75% cup favourite shows about 59%) and cup rows rank and tier on that number.
  • Lagging sources filled, stale ones flagged. When a free results file stops scoring games (openfootball from 20 September 2026) the missing results are read from ESPN under the same spellings; a league still missing results after that is flagged under Celtic's Law rather than priced on a stale table.
  • One order on the Daily List. Confidence first, tested rate second; Group A is the strongest five at 85% or better and is never padded, Group B the next five. On the record, 85%+ list picks land about 90%, 80-85% about 83%.

Tested and left out

  • Recent form beyond a small capped nudge, last season's head-to-head, rating drift, games played, separate home and away ratings, rest days and per-band temperatures all add nothing measurable once the model's own probability is known.
  • Tested since and also left out: a form-led model and form weighted by stage of season (worse on both seasons), team news from injury and suspension lists (nothing that passes), table position, recent goal difference and last-three points (worse at any strength), penalties for weak away sides and habitual losers (the ratings already carry it), a global home-and-away tilt (better, but under the gain bar), damping the FIFA ranking's biggest lifts, a Championship goals scale, and hot form in tennis. Home advantage per league helps a little and is refitted every Monday; it switches itself on only when it clears the gates.
  • The one signal that does add something is bookmaker agreement, about seven times the rest combined. It carries team news the model cannot see, and it is deliberately not used. On the site's own archive the market beats the model on log loss, and picks the model rates five or more points above the market land at the market's rate, not the model's.
  • Every change has to beat the live model on fit, confirm on the current season, and clear gates: 250+ check games, a paired bootstrap with p(worse) at most 0.30, and a minimum gain in log loss. Nothing is ever fitted on this record.

How it is graded

Only prices published before kick-off count. Each morning's predictions are archived as they were published, with the publish time recorded since 2 October. Results are matched back on both team names. A price that appeared after kick-off is never graded, and nothing is ever fitted on this record.

Won and read right. The model almost never picks a draw, because a draw is rarely the single likeliest result. So "won" is strict: the top pick happened. "Read right" adds games where the predicted scoreline was a draw and the game ended level. That is a correct read of the match, so the headline record uses it. Its quoted side adds the draw chance on those games, so a band's quoted and landed figures stay comparable.

Bets are different. On the Bets tab a group is a bet, so a draw still loses a win leg there. The Card, the Edge flag, the slip and the Scale tab each keep their own record, graded at the prices that were available before the start, and each has a pre-registered test it must pass before it is called a method rather than an experiment.

What the priced record says so far (the results ledger, every pick with a pre-start bookmaker price): short favourites hold their value (picks at 1.00-1.15 land 94.7% against 95.1% implied), and the longer the price the more a pick loses to the margin (1.50-1.80: 53.5% landed against 62% implied). Picks where the model sits five or more points above the market land about half the time. An accumulator cannot make an edge its legs do not have; it compounds whatever each leg gives up.

Where the results come from. openfootball fills in results a few days after the whistle. ESPN covers the gap and the competitions openfootball does not. Results are kept once settled, so the record never forgets a game.

What it cannot see

Team news. Injuries, rotation, a new manager or a side with nothing to play for are not in the data. That is most of what a person adds when checking a game by hand, and most of the remaining misses.

Blind spots, flagged rather than hidden. Promoted sides, unrated clubs, cross-division cup ties, clubs with no prior season and leagues whose results are running late are flagged (Celtic's Law) and dropped a confidence tier. They never make the Daily List.

The other sports have their own blind spots. College basketball and college football carry much of last season's rating into a new roster. In testing, home college picks land above their quote and away or neutral-court picks do not, so those are marked Away or Neutral on their rows; college football is off the Daily List on its live record, college basketball is on it at 85% and over. Tennis is on the list at 75% on both tours; the ratings know hard courts best.

Small samples at the top. Calls at 85% and above are rare, so one or two results move a band by ten points or more. Read those rows as direction, not proof.

Roughly one Daily List pick in five or six still misses. That is sport, and a higher bar does not remove it.