Analysis
How the model performs by competition
The same model runs on every league, and the results are wildly uneven. Two competitions beat the baseline comfortably. One manages 23%. Here is the whole table, and what explains the spread.
| Competition | Graded | Accuracy | Draw rate | Home wins | P/L |
|---|---|---|---|---|---|
| Serie A | 10 | 70% | 10% | 30% | −6.12 |
| Premier League | 10 | 60% | 10% | 70% | +11.68 |
| Primeira Liga | 15 | 60% | 13% | 53% | +7.31 |
| La Liga | 18 | 50% | 39% | 39% | −17.15 |
| Eredivisie | 15 | 47% | 20% | 27% | −10.15 |
| Brasileirão | 25 | 44% | 28% | 52% | +12.51 |
| Championship | 24 | 29% | 38% | 33% | −9.58 |
| World Cup qualifiers | 18 | 28% | 28% | 22% | −11.59 |
| Copa Libertadores | 13 | 23% | 46% | 23% | −14.76 |
Competitions with fewer than ten graded picks are left out: Ligue 1 has nine, and the Bundesliga and Champions League have none yet this season. Reporting a percentage on three matches would be noise dressed as a finding.
Draws explain almost everything
Sort the table by draw rate and the accuracy column sorts itself, nearly in reverse. The three best-performing competitions have draw rates of 10%, 10% and 13%. The three worst have 46%, 38% and 28%.
That is not a coincidence, and it is not really a flaw in the model so much as a property of football. A draw is what happens when two sides cancel out. Nothing in a goals-and-results model reliably predicts mutual cancellation, which is why our Pi layer caps its peak draw probability at 0.34 — pushed higher, it starts tipping draws and loses more than it gains.
Copa Libertadores is the extreme case. 46% of graded matches finished level, roughly double a typical league. Two-legged knockout ties reward not losing the first leg, and the competition spans altitudes, travel distances and pitch conditions no goals model can see. Our accuracy there is 23%. We publish it rather than quietly dropping the competition.
Home advantage is not a constant
Compare the Premier League at 70% home wins with the Eredivisie at 27% across our graded sample. The model fits home advantage per competition rather than assuming a single value, which is the only reason the Eredivisie number is not far worse.
World Cup qualifiers get home advantage disabled entirely where the competition is played at neutral venues. Even so it is one of our weakest returns, at 28% — international football mixes squads that rarely play together with wildly mismatched opponents, and a club-form model has little to grip.
The most useful row in the table
Serie A: 70% accuracy, and −6.12 P/L.
Our best tipping accuracy anywhere lost money. That single row is the clearest argument for the distinction the whole site is built on: the tip is the most likely outcome, the bet is the best-priced one, and being right often at short prices does not pay. Brasileirão is the mirror image — 44% accuracy, our best P/L at +12.51, because the value came at longer odds.
If you take one thing from this page, take that. Accuracy is the number prediction sites advertise. It is not the number that pays. The reasoning is here.
What the sample sizes permit
The largest sample here is 25 matches and the smallest is 10. At those sizes a run of three results moves a percentage by ten points or more. Treat every row as a direction of travel, not a measurement.
The overall record — 157 picks, 43.3%, −21.4% ROI — is the only figure with enough behind it to argue about, and it is currently below the always-home baseline.
Figures from the forward log, current as of 27 August 2026. Every pick counted here was published before kick-off and graded against the real result.