Data-driven deep dives across everything we model, from the NBA and MLB to tennis, soccer, and the World Cup.
Every FBS game, every week, published before the market settles. Also published: a 7,606-game walk-forward backtest, an early line value tracker that reads negative in week one, a public pick record that's 15-17, and the two real bugs we found auditing ourselves in August.
Our tennis model flagged 131 bets across the ATP draw in Montreal and finished 73-57 for +22.4 units at flat stakes. Every pick was recorded before the match and graded after. Here's the ledger, what carried it, and the two rows we're not bragging about.
A full 128-runner outright board for both US Open singles draws, weeks before the bracket is made. Our probability, the market's price, and the expected value of every runner the books list — including the one row where the market clearly knows something our model doesn't.
Our World Cup model finished 16-6 on picks with +35.9% ROI on flagged edges — every number published before kickoff and graded in public. On August 21, the same model architecture starts covering the Premier League. Here's the final ledger, what transfers to club soccer, and what we're building next.
20 matches graded, a 15-5 pick record, and every flagged bet published before kickoff — not after. Here's the full ledger, the model behind it, and how Argentina's semifinal win over England added another unit.
Brown called it 'AI hoops.' An ESPN source called him the seventh-best player on a team. Our model has 5,840 possessions of him this season — here's what they say about the trade that sent him to Philadelphia, with no narrative attached.
Thirteen iterations, five seasons of data, and every mistake along the way. This is the honest story of how Predictium's NBA prediction model went from a basic ensemble to a 3-head architecture — and what we learned by getting it wrong first.
