Is the Ghost guessing? What 17,669 settled predictions showed

Measured on 8 October 2026 over 17,669 of the Ghost's predictions on 385 finished matches. The bookmakers' closing prices were the more accurate, and the way the site picked selections from the Ghost's numbers had been choosing its errors. This is what we found, what we changed, and what the study could not show.

The question

The person who runs SoccerGhost looked at the Ghost's daily long-odds bet (the hot bet) and said it felt as if the Ghost was guessing, and wanted facts rather than a guess. We answered with a study, not a reassurance. It took the matches finished between 23 September and 7 October 2026 and, on each selection, the last prediction the Ghost had stored, leaving out any written after kick-off: 17,669 settled predictions on 385 matches, every market. It set each one beside what happened and, where the bookmakers priced it, beside their fair price at the close, with their margin taken out.

What the bands can and cannot show

Grouped by what the Ghost said, with every market pooled together, its predictions came in about as often as it said:

That check is weaker than it looks. Pooling puts an over at 40% in the same band as an under at 40%, whose over was at 60%. If the Ghost expected too few goals, the first would come in more often than it said and the second less often, and the band could still look right. So a table like this can show a forecaster that is too sure of itself, or not sure enough, across the board. It cannot show one that leans one way, such as toward too few goals, and it is not evidence that the Ghost has no such lean. The site's method page says where the Ghost is known to lean: the club model's goal lines run a little low.

The market was more accurate

On the selections the bookmakers priced, their closing prices, with their margin taken out, forecast better than the Ghost: by 0.012 in log loss (a 95% range of +0.004 to +0.021). The gap was clear in the match result, the Asian handicap and team totals. In over/under, both teams to score and club matches there was no clear difference either way.

And wherever the Ghost and the market disagreed, the results sided with the market. Look at the kind of selection the slip builder takes: those the Ghost gave 30% or more, with a price on offer. On those the Ghost rated 10 to 15 points above the market, it said 59.4% on average, the market 47.3%, and 44.1% came in. Where the two were within two points of each other, they said 64.8% and 64.8%, and 64.6% came in.

What was wrong: the choosing

The disagreement mattered most. Anything that chose selections by how far the Ghost rated them above the price was choosing the Ghost's errors, and two parts of the site did. The slip builder looks for the slip the Ghost gives the best chance at a set price, which favours exactly those selections. The confident free picks were ranked by the gap itself. Rebuilt over the same two weeks:

The check is described below. Its two-point line was chosen on these same matches, so the second column is its best case; only matches played since can test it.

The record's 1,625 settled calls told the same story at the time. The headline calls came in about as often as they said: the likeliest result said 57.0% and 59.0% came in; the double chance, 72.4% and 72.5%. The 1,224 value picks made under the rule used before 4 October, which chose by the Ghost's own edge over the price, were over-sure: they said 39.9%, and 33.7% came in.

What changed

After the study, the site offers a selection only where the market has its own view of it and the Ghost is no more than two points above that view. That covers the slips it builds, the daily long-odds bet among them, and the free picks chosen on confidence. Value picks are not held to it: they pass their own test against the market's fair price (how a value pick works). The line sits at two points because of the slip builder's kind of selection, above: within two points of the market, the Ghost said 64.8% and 64.6% came in; two to five points above, it said 61.4% and 58.0% came in.

The check changed what the site offers, not the Ghost's chances. They are still worked out from football alone. A bookmaker's price never moves them; it only decides what the site will offer. Since 25 October the same two-point check also holds back the Ghost's headline result call where it sits further above the market than that, and a pick in another market stands in its place, called low confidence where no other market passes the check either. The Ghost itself has changed since the study. Among other changes, the club model has had four new versions, the latest two live together from 25 October, and the model for the European cups now prices the Conference League too and reads more leagues to rate the cups' clubs. The method page says how it works now.

In the same rebuild, which is the same best case, the free picks with the check were shorter: odds of 1.36 on average, against 2.57. The long-odds bets, at odds of 5.00, were given 21 in 100 on average, close to the 20 in 100 those odds imply, against 43 in 100 as they had been built; 8 of 34 won, against 7 of 36.

What this does not show

Two weeks of matches is not long, and the ranges are wide. The free picks with the check said 78.2% and 73.7% came in, 4.5 points over-sure; resampling the days, a 95% range runs from 6 points under-sure to 16 points over-sure. The study should be run again once a month of results is in. The record keeps counting in the meantime.