Predictions, written down early enough to be wrong

Why this might be worth trusting

Almost nobody writes a football prediction down precisely enough to be wrong. The usual form is a claim vague enough to survive any result, published after enough of the result is known to be safe. This site is an attempt at the opposite discipline, and the discipline is entirely mechanical:

• Published before the deadline. The page for a gameweek goes up while the outcome is still unknown to everyone, including me.
• Hashed and timestamped. Each week records the hash of the exact data snapshot the solver consumed, and the moment it was taken. The same snapshot and the same formula version reproduce the same fifteen names.
• Never edited. Not a word of a published week is changed afterwards. Corrections are appended underneath, dated.
• My disagreements are declared in advance. Where I overrule the model I say so, on the same page, before the games, with the number it costs me.
• Bad weeks stay up, set at the same size and in the same place as good ones. A track record that quietly loses its worst entries is not a track record.

What it does not claim

The model is not good yet. It is version v1.0-phase0: it has no bonus-points term, it matches only 81% of selectable players to prior history, and for 18 of the twenty clubs it prices fixture difficulty from a published rating rather than from expected goals. The method page lists what is known to be broken, in more detail than is flattering.

What is being tested here is narrower and more interesting than “is the model good”. It is whether a mechanical process, honestly published, beats the judgement of the person who wrote it — and by how much, measured rather than argued. Gameweek 1 opens that argument −9.77 expected points in the model’s favour.