Research notes from the Edge Labs desk. How the ratings are built, what the database says, and the work behind the picks.https://www.edgelabs.beten-ushttps://www.edgelabs.bet/research/what-our-model-refuses-to-claimhttps://www.edgelabs.bet/research/what-our-model-refuses-to-claimOur NFL model beats its own predecessor on margin accuracy, loses to the closing spread in every held out season, and we print both numbers instead of picking the one that flatters us.Wed, 09 Sep 2026 00:00:00 GMTnflcfbmodelsmethodologyhttps://www.edgelabs.bet/research/the-preseason-prior-is-worth-eight-gameshttps://www.edgelabs.bet/research/the-preseason-prior-is-worth-eight-gamesEvery team in our college football rating starts the season at its preseason number, then updates one game at a time. The weight on each new result shrinks all season, calibrated so the preseason prior carries the same influence as about eight games of evidence. Here is the math, and the two seasons where we checked it against what actually happened.Mon, 07 Sep 2026 00:00:00 GMTcfbmodelsmethodologyhttps://www.edgelabs.bet/research/why-our-model-demoted-denverhttps://www.edgelabs.bet/research/why-our-model-demoted-denverDenver tied for the best expected win total our first NFL rating produced. In the model that replaced it, the Broncos rank seventh, and the reason is the gap between winning games and winning them by enough.Wed, 02 Sep 2026 00:00:00 GMTnflmodelsmethodologyhttps://www.edgelabs.bet/research/what-the-transfer-portal-is-worthhttps://www.edgelabs.bet/research/what-the-transfer-portal-is-worthOur college football rating carries two portal inputs. The raw count of net transfers came out slightly negative and the average star rating of the incoming class came out positive. Volume is not the signal. Quality is, and here are both weights next to the teams that show the split.Mon, 31 Aug 2026 00:00:00 GMTcfbmodelsmethodologytransfer-portalhttps://www.edgelabs.bet/research/what-a-first-year-head-coach-costshttps://www.edgelabs.bet/research/what-a-first-year-head-coach-costsA first-year head coach costs a team about 2.3 rating points in our college football preseason model, and every one of the six weights behind that number is public.Fri, 28 Aug 2026 00:00:00 GMTcfbmodelsmethodologyhttps://www.edgelabs.bet/research/the-one-totals-angle-that-cleared-our-gatehttps://www.edgelabs.bet/research/the-one-totals-angle-that-cleared-our-gateWe tested dozens of college football angles against eight seasons of closing lines and killed almost all of them in public. One survived: early season unders on high totals. Here is the rule, the evidence, and the live count as the 2026 test grades itself.Fri, 28 Aug 2026 00:00:00 GMTcfbtotalsmethodologyforward-testhttps://www.edgelabs.bet/research/what-returning-production-is-really-worthhttps://www.edgelabs.bet/research/what-returning-production-is-really-worthReturning production is the input everyone quotes in August and almost nobody prices. In our college football model it carries a weight of 5.87 rating points per unit, third behind performance carryover and talent, and here is what that means with the real 2026 numbers behind it.Wed, 26 Aug 2026 00:00:00 GMTcfbmodelsmethodologyhttps://www.edgelabs.bet/research/the-cfb-model-in-seven-numbershttps://www.edgelabs.bet/research/the-cfb-model-in-seven-numbersOur college football preseason rating is one line of arithmetic: a constant and six inputs. It beat an SP plus carryover baseline on average out of sample, across two held out seasons and without losing either, and here is every number, including the one that came out backwards.Mon, 24 Aug 2026 00:00:00 GMTcfbmodelsmethodologyhttps://www.edgelabs.bet/research/how-we-built-our-nfl-power-ratinghttps://www.edgelabs.bet/research/how-we-built-our-nfl-power-ratingWe published a simple model, got called on its blind spot, and rebuilt it on 6,223 games of opponent adjusted margins. The new version won all four test seasons. Here is the whole build, including the number most services would bury.Wed, 19 Aug 2026 00:00:00 GMTnflmodelsmethodology