Abstract Accurate short-term forecasting of survey indices is critical for robust stock assessment. We evaluated whether time-series (TS) models can provide accurate forecasts of Northeast Arctic cod survey indices, and assessed their predictive skill in comparison with the operational Stock Assessment Model (SAM). To ensure a rigorous comparison, both approaches were evaluated using an identical iterative cross-validation scheme, combined with EWMA smoothing to reduce noise in younger age classes with higher variability. Using a moving window approach, the models were refitted annually after adding the previous year’s prediction to the training set. The TS models successfully captured biannual seasonality, autocorrelation, and latent periodic structure in the indices. Results showed that SARIMA-based approaches achieved comparable or higher predictive skill, particularly for younger and more variable age groups, although the SAM maintained better fit. When the same EWMA preprocessing was applied to the assessment inputs, short-term predictive skill also improved. These findings highlight that tailored time-series techniques can provide robust, assumption-light, and data-driven forecasts of survey indices, while also offering a practical framework for evaluating predictive performance and improving assessment models.
Yasaman Maleki (Fri,) studied this question.