Methodology — ATS Edge

The honest headline. Nothing here predicts who wins. The card does something narrower that holds up: it finds a sportsbook whose price is out of line with what the rest of the market says, and tells you which book has it. The edges are small — usually 1.5% to 5% per bet — and many slates have none. A page promising five strong plays every day would be selling you something; this one isn't.

How a price is judged

A dozen books post every game, and they don't all move at the same time. The combined opinion of the market is a better estimate of the true price than any single book's number. So every bet is judged against the other books:

Remove the vig. Each book's two-way price (say −105 / −115) is converted to probabilities and scaled so they add to 100%. That's what the book really thinks.
Put every book on one scale. One book's −2.5 and another's −3 −105 aren't directly comparable, because the half point across 3 is worth far more than a half point across 12. Each book's no-vig price is converted into the average final margin (or total) it implies, using a margin distribution whose spikes at 3, 7 and 10 come from real results.
Take the consensus, leaving the book out. The fair number for a book is the weighted median of everyone else. Sharper books count double, and books that share one price feed (the Kambi group) count once.
Price the bet. The book's actual line and odds are priced against that fair number: chance to win, chance to push, and expected value per dollar.

Prices are only ever combined as probabilities. American odds can't be averaged: −102 and +100 are almost the same price, but their average is −1, which would read as a 100-to-1 payout. An earlier version of this page did exactly that and showed one NFL side at over 5,000% EV.

What makes the card

Sides need at least +1.5% EV against the consensus; totals need +3%, because totals held their value less reliably in testing.
The consensus needs at least five independent books.
A book more than 2.5 points away from everyone else (3 for totals) is treated as a stale or broken feed and ignored. Its “value” would vanish the moment you tried to bet it.
Anything showing more than 12% EV is treated as a data error, not an edge.
By default only regulated US books (DraftKings, FanDuel, BetMGM, Caesars, BetRivers, ESPN BET, Fanatics, Hard Rock and others) can be recommended. Offshore books still help form the consensus.
One wager per game, and no team appears twice, so the card never stacks correlated bets.

Labels follow expected value: LEAN at the minimum, PLAY from +2.5%, STRONG PLAY from +4.5%. Stakes use a quarter of the Kelly criterion, capped at 2 units, so they are small — usually half a unit — because the edges are small.

Every pick shows its fair price: the worst odds at which it is still a bet. If the book has moved past that, skip it.

Why the old model was retired

Until October 2026 this card was picked by an ensemble of a dozen statistical models (gradient-boosted trees, linear models, Elo, power ratings) built on box scores, weather, rest and injuries. It was tested properly: walk-forward, retrained every two weeks, scored only on games it hadn't seen. Against closing lines it found no edge — every confidence interval spanned zero (results below). Live, every one of its 61 graded picks was an NFL side; they went 27–32–2 for −14.7 units, and only 16 got a better closing number than the one published. It routinely claimed 15–40% EV on bets that won less than half the time.

It also explains why college games almost never qualified: the ensemble had learned, correctly, that it knew nothing beyond the college line, so its projections sat within a point of the market and never cleared the bar.

Evidence for the consensus approach

On the 2026 line history (every book, captured every 15 minutes), prices flagged by this method in the morning were re-priced against the consensus just before kickoff. On college spreads, those flagged at +2% or more were still worth about +7% at the close and beat the closing consensus 83% of the time. NFL flags were rarer and smaller but also held up. Totals were weaker, which is why they need a bigger cushion. The sample is still small (dozens of bets, not thousands), so the live record is the real test. That's why every graded pick records its value against the close.

Why closing-line value matters more than this week's record

If you take a price and the market's final, sharpest opinion says it was worth more than you paid, you made a good bet whether it won or lost. Every graded pick here records exactly that: the EV of the price taken, re-priced against the consensus at kickoff.

Win rate over a few dozen wagers is almost pure noise — a 55% record over 40 bets is entirely consistent with having no edge at all. CLV stabilises far faster and is the leading indicator that a process is sound, so it is reported on the performance page whether it flatters the card or not.

Known limitations

The edges are fleeting. A book that's out of line usually corrects within hours. The card is built in the morning; by the time you read it the price may be gone. Check the fair price before betting.
Books limit winning players. Consistently betting the stale side of soft books gets accounts limited. That is a real cost of any approach that works.
No information beyond the market. The card can't see an injury the market hasn't priced yet. It only knows when one book disagrees with the rest.
Thin markets are excluded, not modelled: small-conference college games with few books posting don't reach the five-book minimum.

The retired model's out-of-sample results

Walk-forward, retrained every 14 days, scored only on games each model had not been trained on. Kept here because they are the reason it was retired.

Model history

Data pipeline health

Model weights, feature definitions and source code are not published. Everything above describes the approach at a level intended to let you judge whether the process is sound — which is the part that matters.