Does the model add an edge?
Compare predictions with the market, preserve the misses, and collect new evidence before drawing conclusions.
Local export: . This page is a saved snapshot, not a live feed.
Historical baseline
January–June 2026 has already been examined. Historical candle prices do not prove that a trade could have filled. Actual historical fees remain unverified.
Prediction accuracy
Brier score measures probability error; lower is better.
Where Elo falls behind
Compare probability error as Elo moves further from the market. Large disagreement is not, by itself, evidence of an opportunity.
Descriptive buckets across all usable development and examined evaluation games, including playoffs. These buckets were not used to select a new trading threshold.
Market plus model
One fixed model adjusts the market probability using Elo. It is compared with both the raw market and a separately calibrated market. Development folds use only earlier games for training.
Calibration by probability band
Small bins are noisy. These are descriptive results from examined data.
Rest-feature experiment
Scope and uncertainty
Automatic collection
The local schedule runs while this Mac is awake and its user is logged in. A saved status is not proof that collection is still running.
Usable-data coverage
Game coverage and missing observations
Only queried NBA schedule dates are counted. Missing observations are collection gaps, not no-edge predictions. Preseason and future decision windows are reported separately.
Prospective paper record
Predictions and passes are recorded before outcomes. Hypothetical trade intentions are separate from fills; public snapshots alone do not establish execution.
Coverage and tracking limits
Fresh confirmation
Regular-season games only. Preseason observations cannot qualify. The examined historical period cannot be reused as fresh confirmation. No result here authorizes real-money trading.
Frozen model and review rules
Workspace output and delays
Accepted reports and checks are useful work, not proof of a predictive edge. Applied patches are local deliveries, not deployments. Concurrent task waits overlap.
What information can we test next?
Source snapshots and provenance
Hashes identify the exact local input snapshots used for this export. They do not certify historical publication or fillability.