Does the Ghost learn? What changes with every result, and what we have not seen yet

Every finished match moves the Ghost's team ratings; its club, international, European cup and domestic cup models can each earn a correction to how sure they are, from their own settled record; and where a model's value picks in a market have come in less often than it said, by more than luck allows, it agrees to new ones there less readily. Did its forecasts of the result get better from 21 September to 11 October 2026? Not measurably: on the matches finished in those weeks, the bookmakers' fair prices just before kick-off were closer every week.

The study's figures are as of 11 October 2026; how the Ghost works, and where each model's correction stands, as of 1 November 2026.

What happens after every match

• The ratings move. Every club competition with a match to price is fitted again each morning, and again when a match's line-ups come in. The fit uses three years of results at ninety minutes, with recent matches counting more. Every finished match therefore changes the ratings the next forecasts come from. International sides are rated on eight years of matches, all counting the same.
• The method does not change itself. Its formulas and settings never change on their own, only when we change them, and some of our changes have shipped with a measured gain within what luck allows.
• One correction learns from being wrong. The club model, the international model, the model for the European cups and, since 27 October, the model for domestic cups can each earn one correction to how sure they are of the result. It makes the home win, draw and away win chances sharper or flatter, and every market read from the same forecast moves with them. It is fitted on the older 70% of the model's settled matches, and used only if it beats the uncorrected chances on the newest 30% by more than luck allows. On any run where it does not pass, the model's own chances are published unchanged.
• A new version earns its correction afresh. Since 19 October the correction learns only from the current version's own forecasts, and is not tried before 300 of their matches have settled, so a new version starts its count again at zero. The club model changed to version 1.9 on 19 October and to version 1.11 on 25 October. Before 19 October the correction was a curve, and in the first runs that could try it, from 10 October, the club model turned it down: it made the newest matches slightly worse.
• Where each correction stands, on 1 November. Version 1.11 of the club model had enough settled matches for its correction to be tried, and turned it down again: fitted on the older 405 matches, it made the newest 174 worse, not better (log loss +0.014 a match, ± 0.024). The model for the European cups had 107 settled matches and the model for domestic cups 188, both short of the 300 needed, so neither tried one. The international model has had no match to price since 7 October, and had priced 233 before then, also short of 300. All four publish their own chances unchanged.
• Value picks learn from their own record. Where a model's value picks in a market have come in less often than it said, by more than luck allows, the chance it tests a new pick in that market with is lowered toward that record, the more so the longer the record. Until 19 October any shortfall counted, so bad luck alone could tighten a model that was right; now only a shortfall beyond luck does, tested across every model and market at once. The price: where a model really is over-sure in a market, it is tightened there only once its record shows it beyond luck. This only ever makes the Ghost stricter.
• Drift is flagged, not fixed. A check every morning sets each kind of forecast beside what happened. It flags one only where, across 100 matches or more, the results stray three standard errors or more from what it said. Since 19 October a flag must also pass a test that allows for checking so many kinds at once (Holm's method), so that when every model is right, chance alone raises a flag on no more than about one morning in twenty. The price: a real drift may now have to be larger before it is flagged. On 11 October, before that test, it flagged two: the totals of shots, where over the line was said at 47.6% and came in 55.1% over 519 matches, and the international model's largest edges over the price, said at 25.5% and came in at 16.9% over 213 matches. Nothing changes because of a flag until somebody has studied it.

Has it got better?

Not measurably, by 11 October. This is the Ghost's match-result forecast for every finished match, scored by log loss per match against the bookmakers' fair price just before kick-off on the same matches. A positive number means the market was closer. Each figure is ± one standard error. Every one of these matches was forecast before 19 October, by the models as they stood then.

The market's price was closer every week. The weeks differ by less than their own noise, so three weeks cannot show a trend either way.

Why we publish this

A forecaster that only ever tells you about its wins is selling something. The Ghost's chances are worked out from football alone, never from a bookmaker's price. That makes the comparison with the market a real test, and the market is a hard one to beat. If the Ghost ever does better, the record will show it, just as it shows the calls the Ghost gets wrong.