Improving forecasts of sockeye salmon (Oncorhynchus nerka) with parametric and nonparametric models

Abstract

Accurate forecasts of sockeye salmon (Oncorhynchus nerka) in Bristol Bay, Alaska, play an important role in management and harvesting decisions for this culturally and ecologically vital species. We used a suite of parametric and nonparametric models to assess the frontiers in forecast accuracy of Bristol Bay sockeye salmon possible given currently available data. In retrospective performance testing individual models were capable of reducing pre-season forecast error at the river system level by on average 15% relative to a benchmark model. We used an ensemble modeling approach to produce pre-season forecasts based on historical performance of individual models. This ensemble model reduced river system forecast error by 13% on average in 5 of the 7 evaluated river systems, though it increased forecast error by 39% on average in the remaining 2 systems. We found potential for modest improvements in forecast accuracy across a variety of scales. However, all tested models failed to accurately predict certain periods in the historical salmon return patterns, indicating that further forecast improvements likely depend on novel data rather than more flexible models.