noise
Your inputs come back out undistorted
A common industry shortcut is to "add randomness" — jitter projections, normalize ownership to 100%, hand-tune fudge factors until output looks diverse. Every one of those moves injects noise between your research and your lineups.
- Ownership is never normalized. A real NFL slate sums to ~900% (nine roster spots); rescaling it corrupts every exposure downstream.
- Variance comes from the fitted model, not a dice roll. The spread in outcomes is each player's measured volatility — nothing is jittered.
- Runs are reproducible. Pinned seeds give bit-identical results, so a difference between two runs is a difference in inputs, never in luck.