How we test the model
A replayed season., one week at a time., and For every gameweek it sees only what was known before that deadline.
Published Sun 18 Oct, 07:23 UTC
How we test the model
A replayed season. We put the model in charge of a squad from gameweek 1 and let it play the whole season, one week at a time. For every gameweek it sees only what was known before that deadline: fixtures, form, prices and availability as they stood then. It picks its transfers, captain and chips, and we score it against what actually happened.
How we stop it cheating. The model is trained on earlier seasons and judged on seasons it never saw while training, so it cannot have memorised the answers. We repeat that check with different seasons held out, so the result does not depend on one lucky split.
The results. The same numbers you see in the Track record section.
| Season | Final rank | Managers beaten | Points vs the average manager |
|---|---|---|---|
| 2023-24 | 7,716 | 99.9% | +578 |
| 2024-25 | 40,270 | 99.6% | +544 |
| 2025-26 | 271,084 | 97.9% | +386 |
Rank is the final overall rank read off that season's real standings. Points vs the average manager is the gap to FPL's official average over the whole season.
What replays cannot tell you. Past seasons are not this season. Players change clubs, managers change systems, and a model that was right for three years can still have a bad month. The replays show the method works on average over a long stretch. The live calls on this blog, written before each deadline and stamped with their time, are the real test.
Where to check. Every preview is published before the deadline and every review after the results, with the model data time on each post. Read them in order on the blog index.