Comparison
A factual comparison of the public FPL projection models: whose method is published, whose data is reusable, who publishes a graded record, who is free and who machines can read. No accuracy claims across sites.
Every site here publishes a recommended XI before the deadline. We freeze each one in a public git commit as it is published, then score them all on official Fantasy Premier League points after the gameweek — captain doubled, nothing else. Ours are the two squads we actually field.
| Team | Points | Captain | |
|---|---|---|---|
| evmax — Model XI | 37 | — | official points, autosubs applied |
| evmax — Consensus XI | 59 | — | the crowd's template, same grading |
| Fantasy Football Scout — Scout Picks | 41 | — | XI as published (early) |
| Fantasy Football IQ — IQ AI's squad | 45 | B.Fernandes | XI as published (final) |
| Team | Points | Captain | |
|---|---|---|---|
| evmax — Model XI | 76 | — | official points, autosubs applied |
| evmax — Consensus XI | 83 | — | the crowd's template, same grading |
| Fantasy Football Scout — Scout Picks | 88 | — | XI as published (final) |
| Fantasy Football IQ — IQ AI's squad | 64 | B.Fernandes | XI as published (final) |
| Team | Points | Captain | |
|---|---|---|---|
| evmax — Model XI | 44 | — | official points, autosubs applied |
| evmax — Consensus XI | 31 | — | the crowd's template, same grading |
The analyst's view. Each source's projection for every player is frozen before the deadline and scored after: mean absolute error, lower is better, on the players who played 60+ minutes and on the common intersection of all nonempty sources. Original source coverage is reported separately. The 60+ group is selected after outcomes; it is a descriptive slice. A model earns its keep only by beating the two naive baselines.
| Source | MAE, 60+ min | RMSE, 60+ min | MAE, everyone |
|---|---|---|---|
| evmax | 2.468n=192 | 3.477 | 1.891n=343; source coverage=585 |
| Fantasy Football IQ | 2.371n=192 | 3.245 | 1.786n=343; source coverage=617 |
| FPL ep_next (official) | 2.986n=192 | 3.931 | 2.157n=343; source coverage=617 |
| baseline: season pts/appearance | 2.951n=192 | 3.913 | 2.297n=343; source coverage=358 |
| baseline: last-4 mean | 2.951n=192 | 3.913 | 2.297n=343; source coverage=358 |
| Source | MAE, 60+ min | RMSE, 60+ min | MAE, everyone |
|---|---|---|---|
| evmax | 2.576n=194 | 3.496 | 1.866n=387; source coverage=655 |
| Fantasy Football IQ | 2.491n=194 | 3.393 | 1.83n=387; source coverage=654 |
| FPL ep_next (official) | 2.686n=194 | 3.591 | 1.999n=387; source coverage=655 |
| baseline: season pts/appearance | 2.643n=194 | 3.559 | 2.226n=387; source coverage=387 |
| baseline: last-4 mean | 2.643n=194 | 3.559 | 2.226n=387; source coverage=387 |
| Source | MAE, 60+ min | RMSE, 60+ min | MAE, everyone |
|---|---|---|---|
| evmax | 2.589n=207 | 3.667 | 1.716n=406; source coverage=659 |
| Fantasy Football IQ | 2.639n=207 | 3.656 | 1.752n=406; source coverage=659 |
| FPL ep_next (official) | 2.716n=207 | 3.772 | 1.818n=406; source coverage=659 |
| baseline: season pts/appearance | 2.743n=207 | 3.734 | 2.176n=406; source coverage=406 |
| baseline: last-4 mean | 2.743n=207 | 3.734 | 2.176n=406; source coverage=406 |
The two columns captured since gameweek 1, on identical players. The full benchmark joins this table from gameweek 3.
| GW | evmax MAE | ep_next MAE | players | ahead |
|---|---|---|---|---|
| GW1 | 2.734 | — | 58 | ep_next not captured |
| GW2 | 2.848 | 2.909 | 64 | evmax ahead |
| GW3 | 2.404 | 3.198 | 63 | evmax ahead |
| GW4 | 1.207 | 1.243 | 655 | evmax ahead |
| GW5 | 1.148 | 1.166 | 659 | evmax ahead |
| GW6 | 1.103 | 1.176 | 667 | evmax ahead |
Fantasy Football IQ projections are used under their published licence (free to use with attribution — fantasyfootballiq.app). Only derived metrics appear here; nobody's projections are republished. Sources with restrictive terms are invited in writing instead — the invitation and the answer both get published.
Every model grades itself over a different sample: different gameweeks, a different set of players, sometimes a different definition of error. Side by side, those numbers would look authoritative and prove nothing. The benchmark above exists precisely so the comparison can be made honestly: one sample, one yardstick, frozen in public before kickoff.
Our prospective season experiment tracks registered approaches under shared rules, separately from the historical squads.
Every cell links to its source. Checked 24 August 2026. If a row is wrong or out of date, tell us and we will correct it here.
| Site | Method published | Data reusable | Public graded record | Free to use | Machine-readable |
|---|---|---|---|---|---|
| evmax | Published | CC BY 4.0, per gameweek | Public ledger, every gameweek | Free, no ads, no sign-up | JSON API, MCP server, llms.txt |
| Onside Arena | Private — “the exact recipe stays private” | Projections in a public git repo; no reuse licence stated | Public rolling ledger vs ep_next | Free core, paid tiers | API, llms.txt, MCP (v0.2.0, unchanged since 5 Jun 2026) |
| Solio Analytics | Partly described, model not published | Agent endpoint, refreshed 4-hourly; licence not stated | No public graded ledger | Free tier, membership gating | No-auth data endpoint |
| FPL Review | Private | No open dataset | Occasional articles (“Ultimate Truth”) | Free lite model; solver paid (Patreon from €3.90/mo) | Blocks AI crawlers |
| Fantasy Football Hub | Private | No open dataset | Marketing claims, no public ledger | Paid (Starter £11.99/mo) | No public API |
| Fantasy Football Scout | Private | No open dataset | Marketing claims, no public ledger | Paid membership | No public API |
Onside Arena has far more graded history than we do. They have been publishing a rolling ledger with tens of thousands of graded predictions; ours starts at gameweek 1 of this season and currently holds a single gameweek. A record of one week proves nothing yet, and we would rather say so than let the table imply otherwise.
Their infrastructure is ahead in places — a calibration page and a resolved entity on Wikidata, both of which we lack. Solio's model comes from the person who built the solver toolkit the community runs on. We are the newer, smaller project.
What we do claim is narrower and checkable: nobody else publishes the method, the licensed data, the graded record and free access at once.
Our method is written out on the About page and the numbers behind every claim are downloadable from the data page under CC BY 4.0. Projections are frozen before each deadline and graded afterwards in the ledger, with the raw grading files linked per gameweek. Machines can read all of it through the JSON API and the MCP server.