How a betting model is built and graded on OddsDesk

Here's a model. Outdoor kickoff, sustained wind of 13 miles an hour or more at game time, take the under. Below 13, or no forecast on file, sit out. That is the whole engine. It reads the forecast minutes before the pick, refreshes Thursday and Saturday so the number graded is the last one written before kickoff, and it has been run on every NFL game since the week it was published.

That is what a model is here. Not a spreadsheet. A written rule with a gate in front of it and a side behind it, and a referee who does not work for you.

What the rule can read is the same page of facts for every game. The line and where it moved from. The tickets and the money on each side. Rest days, short weeks, the body-clock kickoff for a West Coast team at one o'clock Eastern. The roof, the surface, the wind. Each quarterback's EPA and completion rate over expected, each line's pressure rate allowed, each pass rush's pressure rate generated, the WR1's share of the targets, the injury report, Sagarin's ratings. 84 named fields, and the rule may use any of them and nothing else. There is no "I just like them this week" field.

The Studio starts you with a sentence that has one blank. Take the under when the wind is 13 mph or more. Fade the road team on short rest. Back the side Sagarin likes by two or more. Fade the team missing its starting QB. Fill the blank and the app counts how many games this week fit before you spend anything. Add a second clause with "and when" if you want it tighter. Three clauses is the ceiling.

The engine ends every read with a side and a confidence, high, medium or low, or it passes. A pass is a result. Nothing is graded on a pass, and the record never pretends one was a win.

Then the part the record depends on. The call is frozen at kickoff at the number it named. Chicago -3 stays Chicago -3 when the live line reaches -4.5. The final score grades it, one unit at the stamped price, and the record keeps the losses. Five graded calls before a record is shown at all, because a 3-0 start is not a track record and the app will not let it pose as one. The leaderboard ranks on the last four working weeks with the lifetime beside it, so a hot fortnight cannot sit on top of a desk with sixty graded calls.

Nothing is graded on the past. Your record starts at zero the day you publish, and only the games you have not seen yet can move it.

Publish under a pen name and the desk is on the marketplace with that record attached. A reader who tails it runs your desk, gets your call, and never sees your prompt.