McConnell’s
Calls
The calls are the model’s. The name is who answers for them. Who writes this →
The weekly slate for pro and college football: a lock, a dog, and the board. The number beside each call is the record of calls like it, because a pick without a track record is just a guess with good typography. There is no total on this sheet: our totals number failed our own test for being a real opinion — its over/under side tracked the posted line, not the teams — so the slot sits empty until we have an opinion worth grading.
How to read this page
The big number is the record, not our confidence. It is how often calls of that same type have won, with the range the true rate plausibly sits in, because the sample is finite and pretending otherwise is how a rate off a couple hundred games gets read as a promise. A bar filling past the halfway tick is the only thing on a row that says the call type has ever beaten the market. Where a row reads “no record yet,” the call type has no record we are willing to publish: the college cells were picked after seeing the seasons that scored them, so they are tracked forward from this season rather than advertised backward.
Disagreement is not a win probability. Each call also states how far our number sits from the market’s, as a percentile against five seasons of calls of the same type: “a bigger gap than 87% of our past calls” means the model and the market have been further apart than this on only 13% of everything we have scored. It is a measure of distance and nothing else. Whether loud disagreement wins is a separate question, and the record beside it answers that one. This number used to be the largest thing on every row, under the word “conviction.” Readers reasonably took it for a win probability, so it is smaller now and says what it means.
The stake is the honest number on the row. It is not set by how much the model likes a call. It is set by how the slot has actually done at that price: a slot whose record clears break-even gets sized off that record, and a slot whose record does not gets the smallest stake we track, 0.25 units, with the reason printed beside it. Most weeks, several calls sit at that minimum. If the stake looks small next to a loud disagreement, believe the stake.
Each slot is a different bet on the model. The lock is the moneyline favorite we back harder than the market does: it wins often and pays little. The dog is the underdog with the biggest gap between our number and the price: it loses more than it wins and is graded in units, not wins. The board is where the model has to beat the closing number, which five seasons of our own backtesting say is close to a coin flip. We publish that sentence on purpose. There used to be a totals slot; we suspended it when testing showed its over/under side followed the posted line rather than the teams, and a pick the model didn’t make has no business being graded under its name.
Every call is generated from predictions frozen into our database before the week starts. Nothing on this page can be edited after kickoff, and the grader writes results into the same files it reads. When the live record earns a bigger sample than the backtest, the live number replaces it, labeled either way.