Research notes from the desk.
Our number against the market, the signals that are holding, and what the tape is telling us — published either way, winners and losers on the same page. Free, no account.
$216,028 lands on one total: BAL/TB Over 8 is today's concentrated bet
The board's biggest pregame campaign is 66% one-sided on a single number, and the pitching trend underneath it points the same way.
The Signal Book Holds: VELO_DROP at 25.6% ROI Over 101 Graded Bets
Two proven signals are still paying out-of-sample while today's board shows eight-for-eight over lean and $906,280 of tracked money.
Model prices BAL +135 at 47.9% — a 12.65% edge against a one-sided market
98% of tracked money on this game's moneyline is on Tampa Bay. Our engine, pregame, has Baltimore roughly a coin flip.
Favourite-side bias fails: 268 bets, 268-0, and only +1.0% ROI
The favourite leg of our bias test won every graded bet and still returned +1.0%. Winning 100% of the time is not an edge if the market charged you for it.
One distribution, three prices: why we simulate outcomes, not markets
A single run distribution prices moneyline, run line and total simultaneously — which is why our signals are internally consistent instead of three separate guesses.
VELO_DROP holds at +22.1% ROI across 113 graded bets
A fastball-velocity decay signal has now settled 113 bets at a 61.9% hit rate — and it got worse as the sample grew, which is exactly what we expect.
VELO_DROP: +19.4% ROI across 110 graded bets
A starter velocity-decay signal has settled 110 wagers at a 60.6% hit rate — our best-documented edge of 14 graded signals.
Live, 11th inning, MIL 3-2 SD: our model likes Under 6.5 at +125 — and we still aren't betting it
A 57.6% model probability against a +125 live price implies 29.6% EV. Our engine declined the wager anyway. Here's why both facts are true.
How our lab is wired: the LLM writes, the engine computes
Every number we publish comes from deterministic code. The language model is allowed to select and explain — never to calculate.
$33,981 hit Toronto at 46c — 96% of one market's flow on one side
Concentration, not size, is the signal: nearly all measured money in Boston-Toronto went one direction, against a market pricing it under even.
Why we simulate a game instead of predicting a score
Our NBA model's total-points error was 15.17 vs the market's 14.36 across 1,070 games. That gap is exactly why we stopped predicting scores.
120 positive-EV prices cleared the model. Two cleared the bar.
Every one of the 120 was a live in-game price. Here is the filter that separated them from the two plays we actually took.
We have MIL at 36.5% to win in San Diego. The market is paying +228.
Our largest live disagreement with a posted price: a 19.83% expected-value gap on Milwaukee's moneyline at San Diego.
NBA 90-99¢ Favorites Are Not Mispriced: Our Bias Hypothesis Fails at n=636
Heavy favorites won 99.7% against a 98.8% market-implied rate — a gap that does not survive its own significance test (p=0.25).
The research is free. The live desk is where it pays.
Tonight's slate, our number against the market, and which signals are firing — updated all day.
Get access ›