How PropStein Works: Model, Gating and Verification
PropStein is a decision terminal for player-prop markets. Instead of selling picks, it publishes what a professional needs to trust a price: the model's fair line, the price boundaries where value exists, the evidence behind each projection, and a dated, falsifiable track record.
The model
Projections use a Negative-Binomial / Gamma mixture (v3.2) over receiving volume: per-player target-share roles (recent window vs. an EWMA baseline), per-target yardage distributions, and a compounding receptions×yards CDF for yardage markets. Each snapshot stores its distribution parameters, so every published number is reproducible from its inputs.
Snapshots, not vibes
Every number on this site comes from a pre-kickoff snapshot with an explicit timestamp, a content hash and a verification state (verified / unverified / stale / blocked). Quotes are captured with source and observation time. Live pricing in the terminal is never served from cache: an unavailable quote is considered safer than a stale one presented as current.
Execution gating
Rows are gated by an explicit qualification policy (edge hurdle and Kelly cap) before they can be actioned. Circuit breakers per market stop automated dispatch when feeds degrade. Availability changes (outages, usage shifts) propagate to impacted teammates as quantified volume deltas, not hunches.
Known defects, published
On our week-1 chronological holdout the model's Brier score was 0.2813 against a naive baseline of 0.25 (skill -12.51%) — i.e. the model currently trails the baseline, and we say so everywhere. Role labels remain provisional. Route-participation data is incomplete for part of the roster. Recalibration work is ongoing; until the skill score turns positive, treat all published probabilities as research, not signal.
FAQ
- Is this betting advice?
- No. PropStein is quantitative research about player-prop prices. Nothing here recommends a wager, and the track-record page exists so you can audit the model yourself.
- Why show losing calibration publicly?
- Because a record you cannot audit is marketing. Publishing the full-board calibration, including weeks where only 2 of 37 decisions beat the close, is what makes the good weeks worth believing.
- How current are the numbers?
- Every page carries its snapshot timestamp and quote status. Lines move constantly; the terminal re-verifies prices before any execution decision and blocks action when feeds are stale.
Model limitations
PropStein publishes model limitations alongside every page. Roles and thresholds are provisional pending chronological holdout validation. In our published week-1 holdout evaluation the model's Brier score trailed the naive baseline; we treat that as an open defect, not a marketing problem, and we publish recalibration results as they land. Snapshots are taken pre-kickoff; quotes age and lines move, so every number here carries a timestamp and a verification state. Route-participation and late-scratch information are not fully modeled. Nothing on this site is betting advice.