Every model choice on CompactStats comes from a study you can read. These are the write-ups, with the real sample sizes and the results that surprised us — including the ideas that failed.
6 October 2026What a promoted team is really worth 465 promoted and relegated sides: their numbers from the division they left point the wrong way. Only the direction of the move predicts.
13 August 2026Early-season league averages lie Six matches into a season a league can look like it has changed character. We measured how much of that is real: almost none of it.
13 August 2026Why four strong picks make a weak bet Four 70% calls together are not a 70% bet. They are a 24% bet — and that gap is where accumulator marketing lives.
27 July 2026Home/away form splits are ~94% noise A 30,582-match study: most home/away over-rate gaps are indistinguishable from coin flips — and models that trust them get worse.
27 July 2026What a 70% call really means 44,728 walk-forward matches: calls rated 67%+ landed 68.5% — calibrated, streaky, and not automatically valuable.
27 July 2026Why an honest model almost never picks a draw 45,556 matches: the highest draw probability our model ever produced was 46.8% — draws are real, calibrated, and almost never the most likely outcome.
27 July 2026Travel, breaks, and the European hangover Long trips cost win probability but not goals; form streaks die over long breaks; and the fabled post-Europe slump barely exists.
The full machinery these studies feed is described on the methodology page.