Methodology
EDIT ME — this is placeholder starter text. Replace with a real description of the model(s) behind the picks on this site before treating it as public-facing. Nothing below should be taken as an accurate description of the actual approach until you've rewritten it.
What this is
This page is meant to explain, in plain language, how picks published on this site are generated: what data feeds the model, what the model actually predicts, and how a "pick" gets selected from that prediction.
What to cover here
- Data sources. Where results, odds, and any auxiliary stats (e.g. expected goals, form) come from, and their known limitations or gaps.
- The model. A plain-language description of the approach (e.g. a Poisson goals model, an Elo rating system) — enough for a reader to understand why a pick was made, without requiring a statistics background.
- Pick selection. What separates a "pick" from a raw model output — what edge or probability threshold is required, and how odds are compared against the model's view.
- Staking. How stakes are sized (flat, proportional to edge, Kelly, etc.) and why.
- Known limitations. Be honest about what the model does not account for, and where it's likely to be weakest. A methodology page that only lists strengths isn't trustworthy.
How the preferred combo is chosen
Each fixture the joint engine prices produces more than one combination. Something has to decide which one gets shown first, and that decision is written down here, before it's ever applied to a real fixture.
The engine already computes two numbers for every combo it prices: the joint fair price, which accounts for the correlation between legs, and the naive price, built by multiplying the single-leg market prices as if the legs were independent. The whole point of the engine is that these two numbers disagree, because the legs aren't independent. The preferred combo for a fixture is the one where they disagree the most.
That gap is measured in probability, not price and not percentage. A joint probability of 8% against a naive 4% is a bigger, more real gap than a joint probability of 92% against a naive 88%, even though the second pair looks identical as a percentage difference and the first pair would look enormous as one. Probability space is where the comparison is honest. Price space compresses at short odds and stretches at long ones, and percentage difference rewards long shots for no better reason than being long shots.
One combo per fixture, not one per day. A daily headline invites picking whichever fixture had the most dramatic gap that day, which quietly turns into a performance metric dressed as a feature. Per-fixture keeps it descriptive: here is where independence would have misled you, on this match, today.
This is not a claim that the preferred combo is a good bet, a value bet, or anything to stake money on. It says nothing about whether the joint price is higher or lower than the naive one, only that the two disagree by more than any other combo on the same fixture. It is the same claim the engine makes everywhere else, just pointed at the single clearest example on the slate.
Change log
Consider keeping a dated log of material changes to the model or staking approach here, so readers can see how the record maps onto what was actually running at the time.
- 2026-08-08 — Documented the preferred-combo selection rule for the joint engine: per fixture, the combo with the largest joint-vs-naive gap in probability space, not price or percentage terms.
EDIT ME — replace this entire file.