NFL Player Props

Direction over calibration

The model isn’t necessarily calibrated: its raw probabilities run more confident than reality, so I don’t read its expected value as a literal return. What it does well is directionality: when it strongly disagrees with the market, it’s usually right about which side of the line the player lands on.

So the model’s EV works as a ranking, not a promise. It prices every QB, RB, WR and TE prop on the board, and only the strongest disagreements make the card: plays where the model’s EV is 40% or more at the DraftKings price one hour before kickoff. Players without enough recent usage history are skipped rather than guessed at.

Verification

Every play gets a second opinion before it’s trusted. Projections are cross-checked against RunTheSims, an independent simulation engine, as a second layer of verification. A pick that holds up under a completely separate simulation earns more confidence. One that doesn’t is a flag to dig into the inputs before betting.

How a projection is made

The engine never starts from the sportsbook line. It simulates each game tens of thousands of times from football data alone, and every player’s stat line comes out of the same simulated game, so the outputs always add up: targets, carries and yards on both sides of the ball.

Each simulation passes through several layers:

1. Game environment: how much offense each team is likely to run, and how it splits between passing and rushing.
2. Shared uncertainty: game scripts that move both teams together, so the tails behave like real football.
3. Personnel: who’s available and in what role as of kickoff, so usage shifts when a starter is out.
4. Opportunity: a joint allocation of each team’s volume across its roster.
5. Efficiency: what each player turns those opportunities into.

The result is a full probability distribution for every stat, not just a point estimate. Each posted line is priced from that distribution.

The technical write-up is in the related project post.