Accuracy

We fitted the model on 5 leagues over the 2023-2025 seasons — 5,256 matches in total — and then tested it on the 2025/26 season, which it had never been shown. Of the 1,752 fixtures in that season, 1,709 were scored. The other 43 were excluded because a promoted club did not yet have enough matches for a reliable rating — and we say so here rather than drop them from the count without comment.

We ran two tests and report both results as they came out. First, a naive benchmark that knows nothing about the clubs and just forecasts home, draw and away at their historical rates: the model was significantly better in 5/5 leagues. Second, calibration — whether events we price at 60% actually occur about 60% of the time: the model passed in 4/5.

Claims like “99% accurate”, “guaranteed winners” and “risk-free tips” are everywhere in football betting. Consider what the first would require: a model that is wrong only once in a hundred calls, week after week, in a game where an early red card or a deflected shot regularly overturns the most likely result. If our own model ever began reporting numbers like that, our calibration check would treat it as a fault to fix, not a result to advertise.

So rather than dismissing such claims outright, take them apart. Which matches were counted, and was that set decided before or after the results were in? What does “accurate” mean in that sentence — the definition is usually missing, and it carries most of the weight. And what outcome would prove the claim false? Advertising copy costs nothing to write, and a claim that can never be wrong is the cheapest copy of all.

The fair test, for us and for everyone else, is whether the claim was able to fail. Ours was: 1,709 predictions on an unseen season, misses included in the totals, and the league that broke our tolerance named on this page.

This is not betting advice. Every probability here is an estimate, and any bet can lose. If gambling no longer feels like entertainment, please read our responsible gambling page. 18+.

• A live hit rate for every competition. Per-competition and per-club records stay hidden until a page has collected 100 resolved predictions. With fewer, one weekend can swing the figure by several points, and printing it would present noise as a track record. A missing accuracy panel on a competition or club page means it has not reached that bar yet.
• A profit chart. Returns from a staking plan reflect the plan as much as the model, and they are trivially easy to present in a flattering light. Calibration and baseline comparisons are much harder to massage, so those are what we publish.

A note on scope: the backtest used the same 5 competitions that TipsterGrid publishes today. The model was tested on precisely the football it is now applied to — not on a broader set chosen to make the results look steadier.

For each settled match in our competitions, we find the first pick we published alongside market prices. Then we set the market’s margin-free probability for the pick at that time against the closing probability right before kick-off. Above zero = the price shortened toward our side.

Window: 2026-10-15 – 2026-10-19; model revision 2026-10-15. All values are in percentage points. The “2+ pt edge” column is restricted to picks where we stood two or more points above the market at the moment of publication. CLV only becomes reliable over a long run, so expect this figure to move as more matches settle.