In-season predictions of daily harvest for lower Kuskokwim River subsistence salmon fisheries

Objective Reliable expectations of outcomes are critical for successfully selecting among candidate actions. Accordingly, we sought to develop models for predicting fishery outcomes from daily drift gill-net openers in the lower Kuskokwim River subsistence salmon fishery in rural western Alaska. Methods We calculated response variables that summarized fishery outcomes for each of 40 daily openers during 2016–2023, including trips per day, total salmon catch per trip, and species composition. We constructed regression models to predict these variables from attributes of the opener (time of season, time of day, duration), in-river run conditions at the time of the opener (as measured by a CPUE index collected daily independent of the harvest fishery), and local weather and water conditions. We used model averaging based on Akaike’s information criterion corrected for small sample size, and we predicted harvest by multiplying predicted response variables. We assessed the reliability of the approach using leave-one-out cross validation. Results Day of the season was a critical explanatory variable for trips per day, catch per trip, and Chinook Salmon Oncorhynchus tshawytscha percent composition but was less important for predicting composition of Chum Salmon O. keta or Sockeye Salmon O. nerka. Fishery-independent total CPUE was not as useful for predicting fishery catch rates as anticipated, but fishery-independent species composition information from the current year was very important in explaining variability for the fishery. Weather and water condition variables were not retained in the final analysis after we found a lack of any meaningful predictive utility. Reliability of harvest predictions varied by species and period of the season; however, cross validation indicated that predictions were of sufficient quality to be useful in decision making (median absolute percent error ranged from 24% to 30% for Chinook Salmon and showed a general lack of directionality). Conclusions We believe that this model could lead to more informed decision making via explicit predictions grounded in past experience. However, some care should be taken to ensure that the model is not used in cases of excessive extrapolation.