McConnell's Calls, Week 1: the slate and the math behind every call

Most football picks content works the same way: someone confident tells you a side, the reasoning is a paragraph of vibes, and last month's misses are nowhere on the page. McConnell's Calls does the opposite. Every week we publish a fixed slate, each call generated from model predictions frozen into our database before the week starts, each one printed next to the track record of calls like it. When a slot's record is mediocre, that number stays on the page. That is the product.

How a call gets made

Our models produce a margin, a total, and a win probability for every game on the board. Those numbers get frozen weekly into a table nothing can quietly edit afterward, then lined up against the market: the spread, the total, and the de-vigged moneyline, which is the market's implied probability once the sportsbook's cut is stripped out.

Here is the part that looks backwards. We do not publish the games where we disagree most with the market. We publish the games where we disagree least.

We measured it. Across roughly five thousand college games, we scored our own error against the closing line's error and sorted by how far apart the two numbers were. On the fifty games where our model disagreed most loudly, our average error was 27.8 points against the market's 12.1. Put that in real-world terms: on exactly the games a normal picks column would lead with, our number was more than twice as wrong as the number anyone can already see for free.

The pattern holds all the way down. Sort every game by the size of the disagreement and our error climbs with it, in every band we can measure with confidence. Below a few points, our model tracks the market almost exactly. Above it, we start paying for the privilege of having an opinion.

So the slate is gated. A college call has to sit within four points of the market, a pro call within three. Anything louder gets dropped no matter how good the story is. This week that removed fifteen of thirty-one eligible college moneylines and eleven of twenty-five pro sides.

That is a strange thing to publish. It means these calls are, by construction, the ones where we have the least to say. We think it is the honest version. A model that beats the closing line is a rare thing and ours does not. What it can do is track the line closely and be right about which side of a near coin flip to take. That is the only claim on this page.

What the numbers mean

Every call carries a disagreement figure, the gap between our number and the market's as a percentile against five seasons of our own calls. Worth being blunt about what it is not. It is not a win probability, not a confidence score, and it does not decide which calls run or how big they are.

Stakes are flat. Every call is a quarter unit. We are not sizing up on the calls we like, because sizing on an unproven opinion makes the record impossible to read: a bad month would look identical whether the picks were wrong or the sizing was.

Each slot carries its own record, and they are not all flattering. The pro total slot has hit 49 percent over ninety graded calls, which is a coin flip. It stays on the page anyway.

The NFL slate

THE LOCK: Detroit ML at -305. Model 73 percent, de-vigged market 72. That point of daylight is the whole call. Best record on the sheet at 76 percent over ninety calls, and also the slot where one upset costs the most.

THE DOG: Carolina ML at +120. Model 48, market 44. Dogs are a units play rather than a win-loss play. This slot runs plus 19 percent ROI over ninety calls while losing more often than it wins, and the second fact is why the first is the number that matters.

THE TOTAL: Under 45.5, Green Bay at Minnesota. Our total is 44.4. This is the coin-flip slot, and one thing this season should settle is whether it deserves to keep existing.

THE BOARD: Cleveland +7.5 (our margin 7.2) and Tampa Bay +3.5 (our margin 3.0). Both inside half a point of the market. Neither is exciting. That is the point.

The college board

College is where the gate does the most work, because the lines run wider and the model has more room to be loud and wrong.

THE DOG: California ML at +100. Model 49 percent against a de-vigged 48. A genuine coin flip at an even price.

There is a second filter on college moneylines, and it removed what would otherwise have been this week's pick. We only take a moneyline where the market's de-vigged probability sits between 15 and 85 percent. Without that band the slot would have gone to a 94 percent favorite, on the grounds that we agreed with the market to within two hundredths of a percent. Correctly selected, completely useless. Laying more than five to one is not a decision.

The rest: Texas -30.5 (our margin 32.3), North Carolina +7.5 (5.5), Auburn -7.0 (9.1), Louisville +6.5 (4.3). All inside the four point gate.

Two games were held rather than called. Sacramento State and North Dakota State are in their first FBS season with no FBS games on record, and the model has nothing to say about a team it has never seen. Held is a real outcome, not a failure. If a week produces fewer than five eligible calls, we publish fewer than five.

What we are tracking

The record updates weekly and grades the calls we published, not a version we wish we had published. Predictions are frozen before the week, the closing line is captured separately, and neither can be edited afterward to make the other look better.

The biggest takeaway is that this is a learning record and not an edge claim. We are not telling you we beat the market, because we measured it carefully and we do not. We are publishing a specific, checkable number every week next to the honest history of numbers like it, and letting a season decide whether any of it is worth anything.

Jesse