Research is separated from hindsight: evaluate the market, generate the model projection, apply frozen criteria, record qualifying projections, then grade the results.
Collect the current player-prop market.
Generate independent player projections.
Apply predefined qualification criteria.
Save projections before results.
Compare projections with actual results.
V12 compares historical player production with the sportsbook market line and identifies discrepancies between its projection and the posted line.
Historical player volatility is used as an uncertainty filter. A projection must satisfy the frozen research criteria before appearing as qualified.
Qualifying projections are recorded before results are known, creating a cleaner forward-testing process.
Qualification is not a guaranteed result or validated win probability. Forward results are used to evaluate the system.
A projection is published as qualified only when it passes both predefined V12 criteria. These thresholds remain unchanged during forward testing.
Historical player volatility acts as an uncertainty filter. Lower volatility identifies observations with historically more stable outcomes.
Learn about volatility →The model projection must differ from the frozen market line by at least five units before the observation can qualify.
Learn about model gap →Qualification identifies lower-uncertainty model situations with a meaningful model-to-market discrepancy. It does not, by itself, establish a validated directional edge or guarantee an outcome.
Model performance is evaluated against both the final result and the frozen market benchmark.
The model's estimated player outcome generated before results are known.
Definition →The sportsbook market number captured before the projection is frozen.
Definition →The difference between the model projection and the frozen market line.
Definition →Historical variation in a player's results, used by V12 as an uncertainty measure.
Definition →Mean absolute error between OddsCrew's frozen projections and actual results. Lower values indicate smaller average projection error.
Definition →Mean absolute error between the frozen market lines and actual results. This provides the benchmark used alongside Model MAE.
Definition →The purpose of forward testing is to keep model development separate from knowledge of future results.
Market lines, model projections, qualification status, and supporting model values are recorded before results are known.
Actual results are added afterward. Frozen projections remain unchanged so historical performance can be evaluated against what was known beforehand.
Beginning with 2026 NFL Week 3, exclusion handling follows a prospective policy established before results.
If a player records no participation, the frozen observation remains in the permanent record but is excluded from normal model-performance calculations.
A documented injury may qualify only when it causes participation to end materially earlier than the player's normal game opportunity. The reason remains documented in the record.
Poor performance, game script, blowouts, coaching decisions, reduced opportunities, weather, and other normal football variance are not grounds for exclusion.
Every frozen observation remains visible in the permanent track record, including its market line, model projection, official result when available, and documented exclusion reason.
Learn about excluded observations →