How it works
What the terminal computes, and what it can’t.
A tool that publishes probabilities owes you its method. This page is the whole of it — including the part most sites leave out, which is what the numbers on this site are not.
Read this first
Every price here was collected from a public source.
Prices, opening numbers and public splits are read from pages that publish them openly, on a twelve-minute cycle, and stored as they were seen. Nothing is bought from a data vendor and nothing is filled in when a source is quiet — a figure that was not published is shown as absent, not as zero.
Where it is known to be weak.
The rating knows which teams have been winning and nothing else. In baseball that is not enough: the market prices the starting pitcher, which is the largest single factor in a game and one this model cannot see, so it systematically likes underdogs the market has good reason to dislike. The backtest on the model page shows that as a loss, and it is left showing.
What would show this wrong.
A backtest and a live record are not the same evidence, and the model page keeps them in separate panels for that reason. Replaying the model over finished games is the weaker of the two: it knows how the season went, it cannot be limited by a book, and it takes prices that may already have gone. Where the two disagree, believe the live ledger — and where the live ledger is negative, that is the answer. It is published either way.
What this is not.
Not a sportsbook, not advice, and not a complete view of the market: only the books our sources quote are on the board, so a better price may exist elsewhere. The model is ours and its record is published in full, wins and losses, precisely so you can decide for yourself whether it is worth anything.
The pipeline
1 · Collection
Every twelve minutes, each league’s board is read from the public pages listed in the source review, with one request per page and an honest user agent. A price is stored only when it differs from the last one stored, so the record is a log of what moved rather than of how often we looked.
2 · The market
Each book’s prices are kept separately, alongside the number it opened at. The consensus is the median across books, taken in probability space because American odds have a hole between −100 and +100. Hold, best price and arbitrage all fall out of those quotes; an arb is only counted between prices on the same number.
3 · The model
A rating per competitor, moved only by games that have finished, weighted by margin and by how much a single game means in that sport. Below six rated games it says nothing at all. Where its probability beats the vig-free market by 2.6 points it fires at a Kelly-scaled stake, capped at quarter Kelly, because full Kelly is a good way to be right and still broke.
4 · The ledger
Games resolve, picks grade, and every headline figure on the model page is computed from that list. A signal row is written the moment it fires, carrying the price available then, and that price is never rewritten — grading only fills in what happened. A bad month shows up as a bad month.
What counts as an edge
Take both sides of a market, convert the prices to probabilities, and they sum to more than 100% — the excess is the book’s hold. Strip it proportionally and you have the market’s honest view. The edge is the gap between that number and the model’s, in percentage points.
A 3-point edge is not a 3% return. It is a small tilt that only pays across hundreds of bets, and only if the estimate is honest — which is what the calibration plot exists to check.
Value against arbitrage
Value is a judgement: the model thinks a price is wrong. It can be wrong itself, and any single bet can lose.
Arbitrage is arithmetic: two books price the same outcome far enough apart that backing both sides returns money whichever way it lands. No opinion is involved — only execution risk, and the fact that books limit accounts that do it often.
The board keeps them in separate columns because they are not the same kind of claim.
Why closing-line value
The closing number is the market’s most informed estimate — everything known about a game is in it. Consistently beating it means you are early to where the market ends up.
CLV stabilises after a few hundred bets; ROI needs thousands. So a model with 0% of picks beating the close is showing an edge long before the profit column could prove it — and a model with flat CLV and a good month is showing luck.
Where the money numbers come from
Each game carries a ticket split and a handle split. Tickets follow the favourite and the famous name; handle follows the number, and leans toward whichever side the market has actually mispriced. That is what makes the splits worth reading rather than decorative — and it is why the money page grades the sharp side against results instead of asserting it works.
What would show the method is wrong
- —The calibration curve bending away from the diagonal — stated probabilities that don't happen at their stated rate.
- —Edge buckets that fail to rise: a 9% edge returning no more than a 3% one means the edge estimate is fiction.
- —Closing-line value sitting at or below zero while ROI stays positive — that combination is variance wearing a suit.
- —An equity curve whose drawdown exceeds what the stake sizing should ever allow.
All four are plotted on the model page rather than described here. That is the point.