Blog — page 3
How prediction models actually work, what betting markets really price, and how to judge anyone's tips — including ours, whose track record is public.
Why our model can't tell you if Arsenal are better than Flamengo
Team ratings look like they're on one scale. They aren't. Every league's ratings are anchored inside that league, which is why cross-league comparison needs data domestic fixtures never provide.
ModelDoes squad market value predict football results?
Does squad market value predict football results? It carries real signal, and quietly re-reads results you have. Where it helps, and where it misleads.
ModelWhy early-season football predictions are the least reliable
August is when every model is at its worst, and most of them don't say so. What actually breaks at the start of a season, and how long it takes to recover.
ModelWhy the Poisson model underrates draws in football
Independent Poisson misprices the scorelines football produces most. Where the assumption breaks, the correction that fixes it, and what our own data can and cannot show.
TransparencySample size in betting: how many bets before a record means anything?
Sample size in betting, explained: the variance math showing how many bets a record needs before its win rate or ROI means anything at all.
Match preview
Barcelona vs Rayo Vallecano: what a clean-sheet streak is worth to a probability
Barcelona have not conceded in La Liga this season, or to Rayo at home in four meetings. Our model still makes both teams to score a 48.4% shot. That is deliberate.
Match preview
Juventus vs Parma: why our model refuses to follow the market to 78%
The market prices Juventus at roughly 78% to beat Parma. Our model says 55.6%. That 22-point gap is what early-season sample size looks like in a probability.
Round-up
La Liga matchday 1: what a round played across twelve days does to a model
La Liga's opening round ran from 15 to 27 August. Our predictions for it were built from wildly different amounts of evidence — and three fixtures have no prediction at all.
ModelMachine learning football predictions: does the algorithm matter?
Machine learning football predictions, assessed honestly: what swapping a Poisson model for gradient boosting actually buys — and what it cannot.
Match preview
Napoli vs Como: a home probability under 40% for a side that finished second
Our model gives Napoli 39.4% at the Maradona and Como 33.2%. It is the flattest 1X2 split in the Serie A round, and the goals line is low too. Here is what produces that.
Match preview
Sevilla vs Atlético: the fixture-list problem our model cannot see
Atlético start a forced run of three straight away games because their stadium is unavailable. Our model has no field for that. Here is what it prices instead.
Match preview
Aston Villa vs Arsenal: when the draw is more likely than the home win
Our model makes Villa Park a 26.3% home win and a 26.7% draw. A Premier League home side ranked third at its own ground is rare — here is what produces it.
Match preview
Coventry vs Hull: when a prediction model has no real favourite
Our model splits this 37.0 / 27.3 / 35.7 — a 1.3-point gap between home and away. The market has it at 29 points. Only one of those is an honest answer.
ModelInternational football predictions: why national teams break the model
Why international football predictions are harder than club ones: ten matches a year, a squad that reassembles each camp, and ratings that don't cross borders.
Round-up
Serie A matchday 1: the round that punished narrow home favourites
Three Serie A openers had our model backing the home side without conviction. All three finished 0-1, 0-2 and 0-1 to the away team. That pattern is worth examining.
TransparencyHow accurate is Forebet? How to check any prediction site yourself
Nobody outside a prediction site can confirm the accuracy it advertises — including ours. Here is the 30-day audit that settles the question for any site you rely on.
Match preview
Leeds vs Brentford: when our confidence score sits below our own probability
Our model makes Leeds 38.4% favourites at Elland Road but publishes 30.7% confidence. That gap is not rounding — it is the model flagging a pick it cannot fully back.
Match preview
Lille vs PSG: a twelve-point gap between our model and the market
The market makes PSG 54% at Lille. Our model says 41.7%. Where a gap that size usually comes from, and which side we would expect to be wrong.
Match preview
Man United vs Ipswich: how little one shock defeat actually moves a model
United lost 2-0 at Hull and our model still makes them 58.8% favourites against Ipswich. Here is the arithmetic that caps how far one result can move a rating.
Round-up
What Premier League matchday 1 told us about pricing promoted clubs
The market made the three promoted clubs near-certain losers. Our model priced them far higher. Two of three won — and the same caution cost us at the Emirates.
TransparencyWhat is a good Brier score in football predictions?
What is a good Brier score? It depends on which formula, which benchmark and how many matches — the arithmetic that turns a raw number into an answer.
ModelThe Dixon-Coles model in football, explained with a worked example
How the Dixon-Coles model fixes the Poisson draw problem: the tau correction on four scorelines, time decay, and a worked example from score matrix to 1X2.
Match preview
Barcelona vs Athletic Club: what a postponed matchday 1 does to a model
Barcelona play matchday 1 on 27 August, after matchday 2. What a postponed opening fixture does to a prediction model is not what you would guess.
ModelCan ChatGPT predict football matches?
Can ChatGPT predict football matches? It will produce a confident answer for any fixture you name. Here is why that answer cannot be scored, and what can.