Since 31 July 2026 every rating on CompactStats is frozen before kick-off into a version-controlled ledger, then graded against the official result. Nothing is edited afterwards. This page grades the most recent 30 days of that ledger: 2,475 frozen ratings for matches from 20 September 2026 to 19 October 2026, of which 2,411 have an official result so far. It rebuilds itself after every data refresh.
| Market | Picks | We said | Landed |
|---|---|---|---|
| Over/under 2.5 goals | 2,194 | 57.9% | 57.0% −0.9 |
| Both teams to score | 2,200 | 56.1% | 56.5% +0.3 |
| Over 3.5 goals (over calls only) | 151 | 49.4% | 49.7% +0.3 |
| Match result (W/D/L) | 2,411 | 49.4% | 49.8% +0.4 |
| ★ High-confidence over 2.5 (67%+) | 159 | 69.6% | 69.8% +0.2 |
We said is the average probability we gave the side we picked. Landed is how often that side happened. A well-calibrated model lands close to what it says — over hundreds of games, not every weekend. Rows where we made no pick (European qualifiers, pre-season cup ties, thin data, conflicting signals) are left out, because we said nothing about them.
Our over/under 2.5 picks, grouped by how confident we were. On a perfectly calibrated model every dot sits on the dashed line.
| Confidence | Picks | We said | Landed |
|---|---|---|---|
| 50–54% | 796 | 52.2% | 51.9% |
| 55–59% | 658 | 56.9% | 59.6% |
| 60–64% | 401 | 61.9% | 57.4% |
| 65–69% | 237 | 66.8% | 63.7% |
| 70% and up | 102 | 72.2% | 63.7% |
Dot size grows with the number of picks. Small groups wobble: twenty picks can miss their mark by ten points through luck alone.
| Side we picked | Picks | We said | Landed |
|---|---|---|---|
| Over 2.5 | 1,388 | 58.3% | 60.2% +1.9 |
| Under 2.5 | 806 | 57.2% | 51.5% −5.7 |
| Both teams to score — yes | 1,653 | 56.7% | 58.3% +1.6 |
| Both teams to score — no | 547 | 54.5% | 50.8% −3.6 |
The small figure after each result is the gap in percentage points: negative means the picks landed less often than we said.
Our under 2.5 calls have landed well short of what they claimed in this window. That is the weakness our own record exposed in early October: ratings in the 30s and 40s were too gloomy about goals, which made under calls look safer than they were. We corrected it on 6 October 2026 — what went wrong and what we changed. Ratings frozen before the correction keep their original numbers, so this gap stays visible until the 30-day window rolls past that date. Then this table will show whether the fix works.
Weeks run Monday to Sunday. One week is a small sample — expect swings of several points either way.
| Week of | O/U 2.5 | BTTS | W/D/L |
|---|---|---|---|
| 14 Sept | 53.6% 220 picks | 61.3% 222 picks | 49.2% 250 picks |
| 21 Sept | 56.9% 246 picks | 53.3% 246 picks | 46.5% 297 picks |
| 28 Sept | 56.4% 220 picks | 61.5% 221 picks | 51.5% 231 picks |
| 5 Oct | 58.6% 669 picks | 56.7% 669 picks | 49.6% 708 picks |
| 12 Oct | 56.8% 833 picks | 54.5% 836 picks | 50.7% 919 picks |
Each cell shows how often our picks landed that week, and how many there were.
Thirty days is a short window. The model behind these ratings is tested on 97,729 matches from 2023 to 2026, replayed using only what was known before each kick-off — the method and the results are on the methodology page, and the individual studies are in research.
One more thing a record cannot tell you: calibrated is not the same as profitable. A 60% call that lands 60% of the time still loses money if the price implies 65%. The edge column on every fixture compares our number with the bookmakers' — and if you bet, please bet responsibly.