Updated Sep 9, 2026 · Week 1 board.
A model is a rule that turns a few numbers about a game into one predicted line. Studio gives you eight slots to write that rule and 56 options to fill them. A free account keeps three saved models. You can open the board and train one without signing in.
The smallest model is three numbers and one subtraction
Take the three highest-projected Philadelphia players on the Week 1 projection board. Jalen Hurts sits at 21.07 points. Saquon Barkley is at 17.66 points and DeVonta Smith at 14.79 points. Add them and Philadelphia comes to 53.52.
Run the same three for the Chargers. Justin Herbert is at 18.97 points, Omarion Hampton at 17.58 points, Ladd McConkey at 13.77 points. That side adds to 50.32. Subtract, and the rule says Philadelphia by 3.2.
That is a model. Three numbers go in, one predicted margin comes out. Nobody would bet this version, because scoring projections for three players are a thin proxy for a team.
The shape is right, though. Studio is that same shape with better inputs, plus a training step that sets the weights for you.
What Studio costs before you make an account
Nothing. The board loads with anonymous access, so you can decide slots and train straight away. The estimate panel prices a training run at $0. It runs on this device or a free Kaggle CPU, never a paid GPU.
Two limits arrive later. Saving keeps three models on a free account. Every model you own counts against that three, whatever state it is in. Publishing a model to the public catalog needs a paid, signed-in account, so your first spread model stays private.
Seven steps from an empty board to a trained model
- Open Studio from the top nav. The board loads with eight slots, zero decided, and 56 of 56 options live.
- Pick the sport. Four buttons sit above the board: NFL, MLB, NBA and NHL.
- Decide the first slot, Target. Choose ATS side pick, which predicts which side of the spread covers a game. ATS means against the spread.
- Fill the other seven. Auto-finish takes the best prior option for every open slot. Collapse best does one slot at a time if you want to watch each choice.
- Press Train model. The button counts the slots still open and fills those first, then runs on your own machine.
- Name it in the Name this model box, then save it to your account.
- Open the saved model and read its six trust checks.
Step 3 earns its keep. On an empty board the page says that deciding Target alone cuts 88% of what is left.
Steps 3 and 4 have a one-click shortcut. The template called ATS classic pre-sets 4 picks. It fills every remaining slot with its best option, so the board arrives fully decided. Step 7 is where most people quit, and it is the step that pays.
Six checks decide whether a model is worth a bet
Those checks live on the saved model card, not on the build board. The trust report grades a trained model on six of them. Each one comes back green, amber or red. The first red check raises a banner naming it, so you never average six numbers into one score.
| Check | Green | Red | What it means for you |
|---|---|---|---|
| Sample size | 1,000 graded picks or more | Under 200 | How much history the score rests on. A small sample makes every number under it noisy. |
| Calibration error | Under 0.05 | Over 0.15 | Whether a 60% call really wins about 60 times in 100. |
| Robustness | Under 0.02 | Over 0.04 | How much your spread record swings depending on which weeks you trained on. |
| Vegas correlation | Under 0.92 | Over 0.97 | How closely your line tracks the book. A red here means you are buying the market price, not making your own. |
| Recent 4w ATS | Above 52% | Below 50% | Your rolling four-week record against the spread. |
| Edge half-life | Over 8 weeks | Under 4 weeks | How long before half your edge is gone and you have to train again. |
Vegas correlation is the check that decides whether you wasted an evening. A model that agrees with the book on every game has no opinion of its own to bet.
Why a 60-game backtest proves nothing
Your first trained model will show a record, and that record is a rumor. Flip a fair coin 60 times and it lands 55% or better about one time in four. Stretch the same test to 200 flips and 55% shows up about one time in eleven. The trust report draws its red line at 200 graded picks.
Studio holds the same line on live results. A published model shows a forward spread record only after 20 graded picks land. Closing-line value waits for graded bets priced against a closing number.
Two honesty laws block the obvious shortcuts. Studio refuses the closing line as an input, because that number lands after the game is priced. The settled result is barred for the same reason.
The Receipts Drawer
The report names its first red check rather than blending six numbers into one grade. A blended score hides which part is broken, and the broken part is the whole story. Sample size is the check to read first, every time.
The desk read on a first model: watch the sample count. Everything else on that card is decoration until it clears 200 graded picks. Train one this week, save it, then leave it alone and check the forward record after 20 picks grade.
If Vegas correlation climbs toward 0.97 as you add inputs, you are drifting toward the book. Change an input then, rather than adding one more. The closing-line value tool and the public track record are the two places to sanity-check what the model claims.
FAQ
Do I have to pay to build a model in Studio? No, the board opens with anonymous access and a training run costs $0 on your own device. A free account keeps three saved models. Paying removes the cap and lets you publish one to the public catalog.
What does ATS mean on the template button? Against the spread. An ATS side pick predicts which side of the posted number covers, rather than who wins the game.
Why did my model get flagged on Vegas correlation? Anything over 0.97 means your line moves almost exactly with the book. You are re-pricing the market rather than disagreeing with it, so there is no edge left to bet.
How many games before my record means something? The trust report stays red under 200 graded picks, turns amber from 200, and only goes green at 1,000. A published model needs 20 graded forward picks before it shows any live spread record at all.
Can I use the closing line as an input? No. Studio refuses it under a named honesty law. The closing line is information from after the moment the model has to make its call.
