KenPom + AI shaping brackets
March Madness bracketters are leaning heavily on KenPom rankings and AI‑powered models to pick upsets and Final Four candidates — consensus brackets are tilting toward Arizona as a title favorite ( ). Bracket strategy continues to favor model ensembles and data viz for edge-seeking entries. (sportingnews.com).
Ken Pomeroy’s live March 18 leaderboard lists Duke as the KenPom No. 1 team (Net Rating +38.85) and Arizona at No. 2 (Net Rating +37.60), with Arizona’s adjusted offensive efficiency at 127.6 and its defensive efficiency at 69.9. (kenpom.com). ESPN’s internal survey found Arizona was the most-popular West-region pick — 50 of 60 ESPN analysts chose Arizona to win the West — while RotoWire’s Gemini-powered simulations reported round-by-round win percentages across 100 runs that put Arizona through the West and into the later rounds. (espn.com) (rotowire.com). Proprietary simulators are also amplifying AI influence: SportsLine’s projection engine ran the full tournament 10,000 times to produce an “optimal” bracket and upset-heavy scenarios, and multiple outlets have fed the bracket to large language and multimodal models for full-bracket output. (sportsline.com) (msn.com). Major outlets and services are blending those model outputs into ensembles and interactive visualizations — BettingPros publishes an expert consensus bracket, TeamRankings/PoolGenius offers an interactive bracket forecaster, and FTN and others expose simulators that let users re-run thousands of Monte‑Carlo simulations. (bettingpros.com) (teamrankings.com) (ftnfantasy.com). Concrete upset calls from AI and model-fed brackets include Gemini putting 12‑seed Northern Iowa over 5‑seed St. John’s and 12‑seed High Point over 5‑seed Wisconsin in first-round scenarios, examples that modelers cite when recommending contrarian picks. (sportingnews.com) (rotowire.com). SportsLine and similar services point to past accuracy claims when pitching data-driven entries — SportsLine says its model “nailed” multiple Sweet 16 teams last season and emphasizes repeated large-scale simulation (10,000 runs) as the method for producing bracket variants users can adopt. (sportsline.com) (cbssports.com).
