Beat the Spreadsheet
Eleven AI models each manage a real ESPN fantasy team for a full season, in football, hockey and basketball. They read the news, set lineups, bid on waivers, trade, and forecast every game. One more team in each league is run by a spreadsheet: I start whoever ESPN projects highest and make every other move by a fixed rule, with no judgment at all. Can any model beat it?
Is running a team the same skill as forecasting?
Ranks across all three leagues. Managing is the share of other weekly scores a team beat; forecasting is the Brier score of its weekly cards. Steep lines are teams that are much better at one job than the other.
Where do the models lose points?
Football, . Each team’s actual lineup is compared with the best lineup its roster could have set in hindsight. Bad luck is the gap the projections themselves would have suffered; the rest is the manager’s. The spreadsheet makes no calls of its own, so everything it loses is bad luck.
Do the models know when they’re guessing?
Every pick made with a confidence, across all three leagues, grouped by the confidence stated. A well-calibrated forecaster’s 70% picks come in about 70% of the time. Dot size is the number of picks; groups with fewer than five are left out.
Do they make things up?
All three leagues. Every claim is checked against ESPN’s box pages and the rosters the model was sent that week. How every number is built.