Home › Methodology
Methodology
The same five rules are applied to every claim on this site. They are the reason most of what gets tested here ends up reported as not working.
1. Walk-forward grading
Every rating, rate and model coefficient used to price a game is computed from games played strictly before it. The state is read, the game is graded, and only then is the result absorbed.
Why it matters: a per-game filter that read the row it was grading once turned a player's record of 27 from 33 into 27 from 27, and overstated one prop market by 40 percentage points. Nothing about the output looked wrong.
2. Matched baselines, never 50%
A claim is compared against what the same teams or players do anyway, over exactly the games the claim fired on. For markets with a posted price, the baseline is what the price itself implied.
Why it matters: the home side wins 56.7% of NRL games. Measured against 50%, most of the competition looks like a trend. Measured against the market favourite's 67%, almost nothing does.
3. Every test counted
The number of comparisons run is published alongside the number that reached significance, with the number expected from noise.
Why it matters: 96 referee tests produced 4 significant results against 4.8 expected — the honest reading is "nothing", and without the denominator the same four look like a discovery.
4. Out-of-sample, or it does not count
A surviving finding is split in time. The first half has to predict the second.
Why it matters: the strongest referee signal in eleven seasons went from 62.1% to 53.2% across the halves of one career, and sits at 41.2% in the current season.
5. Calibration before confidence
A stated probability has to mean what it says. Every probability is passed through a calibration step fitted on earlier games only.
Why it matters: before calibration the model called NBA selections 90% that landed 51% of the time, and scored worse than a predictor that says 50% to everything.
What is never claimed
- A return computed from an assumed price. Where no historical price exists, the rows are marked and no profit figure is given.
- An edge from a baseline that is an identity. "Wins the period" is 50% because one side wins and the other loses — beating it is not an edge.
- A rate without its sample size, or a split below the minimum where a percentage cannot mean anything.
The break-even
At standard −110 pricing you need 52.4%, not 50%. At shorter odds you need more: a bet at 1.20 needs 83.3%. Backing every NRL favourite picks 66.2% of winners and still returns −3.8% on turnover. Picking winners and losing money is the normal case.
Informational only, not betting advice · 18+ · Gamble responsibly.