A projection engine for NBA player props.
AlphaLine is a model that fades inflated NBA player-prop unders. The public piles onto overs, so books shade lines up. The model identifies players whose line sits 1 standard deviation above their own median and who lack the minutes or volume to reach it, then posts those unders before tip-off. No touts, no gut calls, no hand-picks.
Most props content is a person with an opinion. AlphaLine is a model with a track record. It does one thing: find lines that books mispriced against the public flow, then post those plays before the game starts. Everything it posts is timestamped before tip-off and graded the next morning, in public.
How the model works
Five steps run end to end, every slate, on an automated pipeline.
Data
Pulls NBA game logs, rotations, injuries, and live odds. Injury and rotation data refresh every 15 minutes and are trade-aware, so a player who changed teams is mapped to the right roster.
Features
Builds 600+ engineered features per player: rolling windows, opponent-adjusted z-scores, line-efficiency deltas, and head-to-head history against tonight's matchup.
Projection
A walk-forward Ridge regression projects the stat line and a LightGBM classifier ranks it. Trained on the past, scored on the unseen future, so there is no data leakage. Picks are timestamped before tip-off.
Selection
A multi-signal combo score ranks the whole board. The model focuses on unders where the line sits well above the player's own median and the minutes or usage to reach it are not there. Only the highest-edge plays are posted, averaging about 2-3 per night so the board stays disciplined.
Grading
Every posted pick is graded win or lose the next morning and published publicly at the Track Record. Losses are never deleted. The board you see is the board that happened.
Backtest vs live, the honest version
On the held-out 2025-26 season (the model was never trained on it), the model went 231-121, 66.0% win rate, +20.2% ROI at best-of-book price, across 352 picks over 129 slates. That is a real number, and it is also a historical evaluation, which means it is a ceiling, not a promise. The live board is making calls into a future it cannot see.
Live results are tracked publicly from launch. We are not going to paper over the gap between held-out evaluation and live results. The live board is the real test, and that is exactly why every graded pick is public at the Track Record. You should trust what survives daylight, not a number from a spreadsheet.
By stat type
The model focuses on unders. These win rates are from the 2025-26 held-out evaluation.
Books overprice shooting volume when a player runs hot. When the line sits above the player's own median and his usage is down, the under hits at nearly 70%.
Assist lines inflate in favorable matchups. The model finds games where the player lacks the minutes or ball-handling share to reach the stated number.
See it for yourself
The fastest way to judge AlphaLine is to watch the board grade itself for a week. The Discord is free, the Track Record is public, and nothing is locked behind a paywall until you have already seen how it performs. The 2025-26 held-out season showed 66% wins and +20.2% ROI. That is what you are betting on.