Updated Sep 9, 2026 · Studio, signed out.
Two numbers decide whether a trained model has anything worth betting: 0.92 and 0.97. Studio's trust report measures how closely your line tracks the book's number. Under 0.92 the check is green. Over 0.97 it is red.
What the check is really asking
A model makes money by disagreeing with the market at the right moments. Land on the book's number every week and nothing is left to bet, however clever the inside of the thing is.
The report turns that idea into one figure. The green band, under 0.92, means your calls show real daylight from the market. The red band, over 0.97, gets a blunt verdict: you are buying the line, not beating it.
The band between them is the honest middle. Studio calls a reading there high, and stops short of calling it broken. That is a model with an opinion on some games and none on the rest.
A red reading can hide under a red line above it
The report grades six checks and raises one banner. The banner names the first red check in order, and correlation sits fourth, behind sample size, calibration error and robustness.
A model with fewer than 200 graded picks turns the sample check red, and that check takes the banner. Your correlation problem is still on the card, four lines down, with nothing pointing at it.
So read all six every time. The banner is a shortcut, and shortcuts skip things.
One input carries the book's own number
The third slot, Features, holds the data a model reads. Fourteen options sit in that catalog, and two of them are baseball-only, so an NFL board cannot reach them.
Exactly one of the fourteen feeds market prices. Here is what the shortlist looks like when you sort it by how close each input sits to the book.
| Features option | What it feeds the model | Carries the book's number |
|---|---|---|
| Market price feed | Prices captured before kickoff, never a post-close price | Yes |
| Public sentiment feed | Sentiment mined from podcast and video commentary | No |
| Crowd ELO ratings | Ratings built from reader votes on one matchup against another | No |
| Team recent form | The recent-performance trend for a team | No |
| Pace and total environment | Team pace and scoring-environment signals | No |
| Weather conditions | Game-time weather | No |
Only the first line quotes a price. The other thirteen options feed team, player and environment signals, and none quotes a number the book has published.
Why the reading drifts upward on its own
A closing price is the market's final and best-informed estimate of a game. Every sharp opinion, every injury update and every large bet has been folded into it by kickoff.
So a build made from good public inputs will land near that price on most games, and it should. The correlation check is not asking you to be different for the sake of it. It asks whether anything is left over once you have agreed with the obvious.
That is why the green line sits at 0.92 rather than at zero. A build inside the green band is not a contrarian machine. It agrees on the easy games and parts company on a handful, and the handful is where a season gets decided.
The law that closes the obvious shortcut
Even the market feed cannot hand you the answer. Studio limits it to prices captured before kickoff, and never a post-close price.
A named honesty law makes that a rule rather than a habit. For a pregame target, the closing line counts as future data, because it arrives after the moment a build has to make its call. Settlement data is blocked by a second law on the same logic.
Both laws are printed on the board while you build, not buried in a help page.
What to change when the reading goes red
Take the market feed out and train again. If the number drops into the green band, market prices were doing your predicting for you.
Then put inputs back one at a time and read the number after each run. Four changes in one run tell you the reading moved without telling you which change moved it.
If the reading stays high with the market feed gone, the drift is coming from somewhere quieter. Take the remaining inputs out one at a time before you start rebuilding the middle of the thing.
Watch the direction rather than a single figure. A model that sits under 0.92 today and creeps past it after two retrains is walking toward the book, and the answer is a different input, not one more input.
The Receipts Drawer
This is the check that decides whether an evening of building produced anything. Over 0.97, a model is a slow and expensive way to copy a number anybody can read for free.
The banner rule matters as much as the threshold does. It names one red check, so a red fourth line under a red first line never gets announced. Read the fourth line yourself.
One desk suspicion, offered as suspicion. Reader votes and podcast sentiment never quote a price, and both track the public mood a book prices in anyway, so a board leaning on them may drift market-ward with no market data in it. Studio does not measure that, so treat it as a thing to test rather than a finding.
Watch the reading across retrains rather than once. If it climbs toward 0.97 while your record stays flat, change an input before you change anything else. The public track record is where a published model's live claims have to land.
FAQ
What is Vegas correlation in Studio? It scores how closely a trained model's numbers track the book's number. Green is under 0.92, amber runs to 0.97, and red is anything above that. Red comes with a verdict attached: you are buying the line rather than beating it.
Why is a high reading bad when my record looks fine? A model that agrees with the book on every game has no disagreement left to bet. Over 0.97 the record you are looking at belongs to the market price, and you are paying juice to copy a number that is free to read.
Which input pulls a model toward the market? Market price feed, one of the fourteen options in the Features slot. The rest feed team, player and environment signals: form, injuries, weather, travel, pace, quarterback ratings, snap usage, scoring splits, matchups, ballpark effects, pitcher mix, reader votes and sentiment.
Can I use the closing line to fix the reading? No. A named honesty law blocks the closing line for a pregame target, because that price lands after the moment a build has to make its call. Settlement data is barred under a second law.
My banner says sample size, not correlation. Why? The banner names the first red check only, and sample size sits above correlation in the order. A model under 200 graded picks takes the banner with it, so a red correlation goes unannounced until you read the fourth line yourself.
