Originally published September 9, 2026 by Model Desk.
Adding player fantasy projections will not tell you how many points a game will score. Build a National Football League (NFL) totals model around the combined scoreboard total instead. In Studio Workbench, choose Total over/under as the target and decide which information the model uses.
Choose NFL and Total over/under
This Studio training workflow currently supports the National Football League (NFL). Training for the National Basketball Association (NBA), National Hockey League (NHL) and Major League Baseball (MLB) is unavailable.
Choose Total over/under under Bet type. The target compares the combined scoreboard total with the posted total. A tied total is a push: the bet ties. The trainer leaves those games out of its win-or-loss training labels.
Set up the totals recipe in Workbench
- Open Studio Workbench and select NFL.
- Under Bet type, choose Total over/under.
- Open Which games to review the games included. Narrow filters leave fewer games to learn from. Start with a broad set of games. Try narrower filters later.
- Open Ingredients to choose the inputs, or features, the model sees. You can select several. Review their roles below before accepting suggested choices.
- Review Method and How it is tested. Probability tuning adjusts the model's win chances. Select Choose suggested option to accept the next recommended choice. Auto-finish fills the remaining open choices; review the completed recipe before running it.
- Press Train model when it is available. If the app reports unsupported choices or too few usable games, address that message before retrying. A selected recipe is setup, not a completed result.
- If training finishes, read the result on past test games and its warnings. Check the number of usable games and which seasons were tested before interpreting a rate.
Workbench includes Calibrated ensemble · sealed, a starting recipe that selects the totals target. Review its choices before training. Applying a template sets up a recipe; it does not produce a trained result.
Evaluate ingredients by what they supply
Team recent form uses past performance. Travel and rest uses rest and division information. These are choices to evaluate, not a ranking of which inputs matter most.
Weather conditions currently uses roof type, not temperature or wind. Weather severity bands is not a supported training filter. Do not build this recipe around the old Pace and total environment option; it is outside the supported training choices.
Market price feed supplies market lines and a game-winner probability derived from prices. Including it does not tell you how often a trained model will agree with the betting line. Compare results on held-out games: games excluded from training.
Use information available before the prediction. Results after the game cannot serve as pregame inputs. Check when market quotes were recorded; the option name does not tell you when the prices were recorded.
Fantasy points are not scoreboard totals
Fantasy points reward player actions under a contest's scoring rules. Scoreboard points measure game scoring. Adding player fantasy projections does not turn them into a forecast of the combined final score. A totals model needs the game-total target, not a sum from a fantasy board.
Read the totals result without a universal cutoff
A sample is the games evaluated. Read its size alongside the seasons, filters and warnings. A larger sample can still reflect a narrow set of conditions. A warning cutoff does not prove that a model is reliable.
ATS means against the spread. The current totals read-out uses Past total accuracy (test seasons). Do not read an ATS record as a totals record. Check the selected bet type and the games behind the result.
The Receipts Drawer
Keep the bet type and test games fixed. Change an ingredient, then compare the results if both builds finish. This helps you see what changed in that comparison. Past test results do not establish how a model will perform on future games.
FAQ
Which Studio target builds a totals model? Total over/under, under Bet type in Workbench. It models whether the combined scoreboard total finishes above or below the posted total. Choose NFL for this workflow.
Can I build a totals model for basketball or hockey? No. This Studio training workflow currently supports the National Football League (NFL). Training for the National Basketball Association (NBA), National Hockey League (NHL) and Major League Baseball (MLB) is unavailable.
Is there a one-click template for totals? Workbench includes Calibrated ensemble · sealed, a starting recipe that selects the totals target. Review its choices before training. Applying a template sets up a recipe; it does not produce a trained result.
Which inputs actually matter for a total? Choose inputs in Ingredients. Team recent form uses past performance; Travel and rest uses rest and division information. Weather conditions currently uses roof type, not temperature or wind. Evaluate the choices on held-out games: games excluded from training. Do not assume a feature improves predictions.
Why does my totals model still show an ATS line? ATS means against the spread. The current totals read-out uses Past total accuracy (test seasons). Do not read an ATS record as a totals record. Check the selected bet type and the games behind the result.
Where these numbers come from
The publication date comes from this article's retained publication record. For the controls and starting recipes covered here, use Studio Workbench and the Studio recipe guide.
