NHL Measured results
No demonstrated edge yet. Model predictions and price shopping are separate experiments.
Model predictions · one saved forecast per game
The first eligible published forecast in the final 72 hours is saved with its exact bookmaker line and prices, at least two minutes before scheduled start. Results use that record only. Missing prices are excluded, never filled with assumed odds.
Prospective collection is not enabled. No verified record is available yet.
Paper returns risk one unit at the recorded price; no execution, stake limits or slippage are assumed. Ties/pushes return the stake; equal forecasts make no directional pick. Negative log-loss difference favors the model. ESPN provides an observation time, not a bookmaker update timestamp.
Research loop · fixed paper screens
Every 15 minutes, collection and grading feed three declared screens across five sports and three markets: published picks, raw model value, and the frozen EV rule. Each producer version stays separate. Rules are registered before their games are captured.
Research registration waits for prospective collection to be enabled.
These are paper screens, not bet recommendations. Returns subtract an assumed 1% cost per unit. Review requires the fixed 56-day horizon, at least 100 settled games across 14 dates, positive early and late results, and adjusted intervals favoring both returns and predictive loss. Approximate intervals account for declared comparisons; they do not certify fills or profit. A new untouched confirmation sample and independent review are still required. Nothing promotes automatically.
Price shopping · no model involved
Each book, side and exact line is recorded once. Closing value requires a separate, later quote in the final hour before start. The flagged book is excluded from consensus; uncertainty is grouped by game.
No verified closing-price sample yet.
CLV is a probability-point difference, not profit. Game lines and props are reported separately. Repeated books on one game do not create independent evidence. Small samples and multiple comparisons do not establish an edge.
Each published model version and source commit is measured separately. Unknown producer records are descriptive only and cannot justify a model promotion. Historical results were recalculated after games, so they cannot prove pregame performance. They are excluded from this record. Model parameters and shopping thresholds remain frozen.