78.5% ML pick claim
- Posters describe an NBA ML pipeline using nba_api, Polymarket and DraftKings data with XGBoost ensembles. (x.com) - The claimed performance figure for that approach is a 78.5% ML picks hit rate across recent samples. (x.com) - The thread highlights features like eFG%, turnovers, ELO and fatigue as key model inputs. (x.com)
A group of posters on X says its NBA winner-picking model hit 78.5% on recent moneyline picks, using game stats, sportsbook odds and prediction-market prices. (x.com) The posts describe a pipeline built with `nba_api` for NBA data, DraftKings lines for sportsbook prices, and Polymarket feeds for market-implied probabilities. The model itself is described as an XGBoost ensemble, a method that combines many decision trees into one forecast. (github.com) (sportsbook.draftkings.com) (docs.polymarket.com) (xgboost.readthedocs.io) A moneyline bet is the simplest NBA wager: pick the team that wins the game, regardless of margin. DraftKings lists NBA moneylines daily, and Polymarket’s documentation says developers can pull live market data through its API. (sportsbook.draftkings.com) (docs.polymarket.com) (docs.polymarket.us) The feature list in the X thread includes effective field goal percentage, turnovers, Elo ratings and fatigue. Effective field goal percentage adjusts shooting to give extra weight to 3-pointers, while fatigue usually means rest disadvantage, back-to-backs or travel load. (x.com) (basketball-reference.com) (pmc.ncbi.nlm.nih.gov) Those inputs are common in basketball models because they try to capture different parts of the same game. Shooting efficiency measures how well a team turns shots into points, turnovers measure lost possessions, Elo estimates team strength, and fatigue tracks schedule strain. (basketball-reference.com) (pmc.ncbi.nlm.nih.gov) (xgboost.readthedocs.io) The 78.5% figure, as presented on X, is a hit rate claim rather than a full audited betting record. The posts visible through search do not provide a public backtest file, sample size breakdown, closing-line comparison or profit-and-loss statement alongside that number. (x.com) That missing context matters because win rate alone does not show whether a betting model beat the market. A model can pick heavy favorites, post a high hit rate, and still lose money if the odds are too expensive. (sportsbook.draftkings.com) XGBoost is widely used for this kind of classification problem because it can weigh many interacting variables at once. In practice, a model can learn that a small rest edge matters more for one team profile than another, or that turnover risk matters more against certain opponents. (xgboost.readthedocs.io) The public materials do support the basic architecture the posters describe: NBA data can be pulled with `nba_api`, DraftKings publishes NBA lines on its sportsbook pages, and Polymarket offers market-data access for developers. The part that remains unverified from public evidence is whether the recent 78.5% run holds up over a larger, fully documented sample. (github.com) (sportsbook.draftkings.com) (docs.polymarket.com)
