Diagnosing stock-assessment misspecification with machine learning
Scientific question
Can the multivariate diagnostic signature of a stock assessment identify the process responsible for structural misspecification, quantify uncertainty in that diagnosis, and recognize cases that are not identifiable from the available diagnostics?
Experimental framework
The core study uses known simulated truth. Stock Synthesis operating models generate pseudo-data, correctly specified and deliberately misspecified estimation models are fitted, and the resulting multivariate diagnostics become features for transparent statistical and machine-learning classifiers.
Initial mechanisms include natural mortality, selectivity, somatic growth, recruitment dynamics, catchability, observation error, ageing error, and catch error. Validation is scenario-aware rather than based on random replicate-level splitting.
Current stage
The workflow architecture is established. Current development focuses on auditing the published sardine-like, flatfish-like, and cod-like operating-model configurations and certifying the assumptions that define simulated truth before controlled misspecification experiments begin.
Planned output
A peer-reviewed methodological paper on probabilistic diagnosis of stock-assessment misspecification, with explicit tests of leave-scenario-out, leave-severity-out, leave-life-history-out, and multiple-misspecification generalization.
The analytical repository remains private during active methodological development. This page is the public project record; it does not duplicate code, model files, simulation outputs, or manuscript materials.