AI NBA models claim 78% win rate
- Moddy.ai said on March 23 its NBA moneyline model “Closeout Rater” hit 77.6% on 49 settled picks, but returned only 1.2% profit. - The company said the model went 8-for-8 over two days, while warning that NBA favorites at -400 odds need 80% wins to break even. - The claim lands in a crowded market of AI sports models, where published academic NBA accuracy ranges still cluster far lower. (nature.com)
A 78% NBA win rate is real in one recent public claim, but the more important number was 1.2% return on investment. (moddy.ai) Moddy.ai said on March 23, 2026 that its “Closeout Rater” NBA moneyline model was hitting 77.6% across its first 49 settled picks. The same post said the model had gone 8-for-8 over the prior two days. (moddy.ai) The company also said that run produced only 1.2% return on investment, because many winning picks were heavy favorites with small payouts. It used a -400 line as its example, saying a bettor needs to win 80% of those bets just to break even. (moddy.ai) That gets at the basic math behind sports models. Picking the team most likely to win is not the same as finding a bet where the posted odds are wrong. (moddy.ai) In NBA betting, favorites win often enough that raw accuracy can look impressive even when the edge is thin. Moddy.ai said its model only “fires” when its estimated probability is above the market’s implied probability. (moddy.ai) That distinction matters because NBA markets are already saturated with data. Pickswise says its supercomputer simulates games 10,000 times and updates probabilities in real time as team news and betting lines change. (pickswise.com) PredixSport says its NBA system is trained on about 40,000 matches and uses hundreds of team- and player-level features, including injuries and projected starting fives. Its public game pages show win probabilities, spreads and totals that update by date. (predixsport.com) A separate blockchain-linked version of the pitch is also circulating. Bittensor lists Sportstensor as Subnet 41, a decentralized sports prediction network with real-time scoring and rewards tied to forecast accuracy. (bittensor.ai) (pm.wiki) But published research still shows a lower baseline than many marketing claims. A 2023 paper in *Symmetry* reported average NBA prediction accuracy around 66%, with a maximum around 78%, while a 2025 *Scientific Reports* paper described ensemble methods as practical without claiming anything close to guaranteed betting profit. (mdpi.com) (nature.com) So the current story is less “AI cracked the NBA” than “AI vendors are publishing sharper probabilities into an already efficient market.” A model can win a lot of picks and still leave almost no margin after the price is baked in. (moddy.ai)
