Abstract
<title>Abstract</title> <p>Distributional forecasting systems often select the model with the best validation score and treat its implied action as resolved. Finite validation data, however, may rank models without excluding other distributions that imply materially different actions. We introduce the Finite-Evidence Decision Identification Procedure (FEDIP), a modular audit that applies a declared competitive-model screen, projects retained distributions into a prespecified action space, and reports the unresolved action range. “Decision identification” is procedural: it means contraction of candidate-implied actions under the declared system, not structural identification. Validation length is supplied evidence rather than an estimated optimal budget. An oracle concentration benchmark shows that evidence demand rises with score noise and candidate-space size and may be unbounded when consequential alternatives are nonseparable under the declared score margin. A Monte Carlo factorial conducted under a locally frozen protocol covers 120 conditions and 48,000 replication-condition runs. Expanding from five to nine models increases the standardized 5% value-at-risk range more at 20 than at 500 validation observations: the stable-process interaction is 0.0526 (95% Monte Carlo interval 0.0440–0.0612), concentrated in mixture and skewed processes. In 2,710 rolling blocks for 11 assets, the screen retains 8.025–8.976 of nine models on average, indicating weak discrimination. A maximum-threshold policy reduces violations from 5.06% to 3.99% but raises the standardized threshold by 9.47% relative to the selected winner and worsens asset-averaged symmetric calibration error. FEDIP therefore diagnoses unresolved action dispersion; final action choice requires an explicit economic loss function.</p>