Guide

How to read the tool (and how not to)

By Adam · Updated 16 July 2026 · 6 min read

The tool outputs numbers, and numbers invite misreading. A 62% that loses is not a lie. A +12% edge resting on nine matches is not a gift. Every pick on the board carries four numbers — a probability, a price, an edge, and a confidence grade — and they deserve your attention in exactly that order. Read them edge-first, the way the eye wants to, and the board will cost you money. Here's how each one is built, what it can tell you, and the specific way each one gets misread.

The probability comes first

Everything starts with the model's estimate of how often the bet lands: recency-weighted rates, shrunk for small samples, adjusted for opponent and referee, then recalibrated against seasons of history. The methodology page shows the working and the modelling guides walk the machinery, so here's the only thing that matters at reading time: a probability is a rate, not a prediction. A 62% means that in a long run of identical situations, roughly 62 in 100 land — which is another way of saying 38 don't. The model is graded on exactly this claim, in public, on the track record: do our 30%s happen about 30% of the time? That's the promise. Single results were never part of it.

The price turns an opinion into a bet

On its own, a probability is trivia. It becomes a bet the moment it meets a price, because a price is just the market's own probability in another form — decimal odds of 2.00 imply 50%, minus the margin the bookmaker folded in. The tool strips that margin out properly where it can, so you're comparing our estimate with the market's honest one. No number on the board means anything until this comparison happens. A brilliant probability at a terrible price is a pass; a modest one at a generous price is a bet.

Edge, and the words the tool uses

Edge is the comparison, compressed: our probability × the price − 1. It's your expected profit per pound staked, if our number is right — and that "if" is the entire game. The board translates edge into plain verdicts. Around 8% and up reads as value; roughly 3–8% is thin; anything near zero is a pass; negative is an avoid. The 3% floor isn't decoration — below it, the edge is smaller than the model's own honest uncertainty, so what looks like a small edge is usually just noise. One design rule worth knowing: on this board, colour means edge and nothing else. Green never means "safe". It means the price is bigger than our number.

Confidence is the weight, not the verdict

Beside every pick sits a confidence grade, and it answers a different question from the edge. Edge asks "is the price wrong?"; confidence asks "how sure are we of the number that says so?" It's built from sample size — the board calls 25+ matches solid, around a dozen fair, and less than that thin — plus, for cards, how mature the referee's own record is, and whether the lineup is actually confirmed. An unconfirmed team sheet discounts confidence heavily, because the quietest way to lose a player bet is for the player to be on the bench. The reading rule: a big edge with low confidence isn't a bigger opportunity, it's the model saying interesting — prove it. The same 62% deserves entirely different money at nine matches and at ninety.

The Why panel is the audit trail

Tap any pick and it shows its working: the per-90 rate underneath, the opponent adjustment, the referee's card tendency (flagged when he's a genuine lift), the derby uplift where one applies, the recent match-by-match counts, and the sample it all rests on. This isn't decoration either — it's there so you can disagree. If a probability surprises you, the panel tells you which ingredient produced it, and then you can argue with that ingredient instead of with a number. A tool you can't argue with is a tipster with better fonts.

Kelly, quartered

Each value pick carries a suggested stake as a fraction of bankroll — a quarter of the Kelly optimum. The bankroll guide makes the full case; the short version is that full Kelly maximises growth at the cost of violent swings and total faith in your own numbers, and quarter Kelly keeps most of the growth for a fraction of the pain. Read it as a ceiling and a discipline, not an instruction. It already scales with the edge and the odds, which means it's also quietly telling you when a bet deserves almost nothing.

The builder's one big trap

The builder multiplies your selected legs into a combined price, and the multiplication is only honest when the legs can't influence each other. Legs from different matches are independent — multiply away. Legs from the same match are not: the game state that delivers one often delivers the other, and bookmakers price that correlation in, which is why the real same-match price is shorter than naive multiplication suggests. The tray corrects for this where it can — it measures how often the legs have actually landed together in the players' own match histories, shows the adjusted true chance and fair price beside the naive figure, and flags redundant legs that add margin without adding value. The bet builder guide covers the full argument.

How not to read it

The failure patterns are reliable enough to list. Sorting by edge and betting the top of the list, ignoring the confidence column that was the whole point. Treating a 70% as destiny and the model as broken the three times in ten it isn't. Betting every green light because green feels like permission — the board some weeks offers genuine value on nothing, and taking no bet is a result, not a failure. And building same-match accumulators off naive multiplication because the combined number looked pretty. The board is a list of arguments, ranked by strength. Which ones you take, and at what size, was always the part no model can do for you.

Frequently asked

Keep going