NBA Player Props: Where AI Prop Predictions Find Their Edge
Prop markets are thinner and slower to correct than game markets. That structural lag is the single largest source of measured edge in the model.
Props are inefficient for structural reasons
Game lines attract the most money and the sharpest attention, so they correct quickly. Player prop markets are posted later, moved less often, and limited more aggressively — which is exactly why a disciplined model finds more mispricing there.
Ascend publishes three player prop over/unders every day alongside the game card, and prop selections are modelled with the same pipeline rather than a simplified shortcut.
The four levers behind a prop projection
Usage rate tells the model how much of the offence runs through the player when he is on the floor. Minutes projection converts that share into volume. Matchup defence adjusts for the specific opponent and the likely defensive assignment. Pace scales the whole thing by how many possessions the game is projected to contain.
Combine the four and you get a projected distribution, not a point estimate. The over/under decision is made against the full distribution, which is why a projection close to the posted line can still carry a real edge if the distribution is skewed.
Discipline on prop staking
Prop variance is higher than game variance, so stake sizing matters more. Every prop selection is capped under the same quarter-Kelly rule, and the published card shows the modelled probability, the de-vigged market probability, the edge and the unit size before the event starts.
Nothing is edited or removed after settlement. The rolling hit rate and return on turnover you see are computed on the full published record.