Methods for assessing and responding to bias and uncertainty in U.S. West Coast salmon abundance forecasts

Abstract

We quantified the bias and accuracy of U.S. West Coast Chinook and coho salmon abundance forecasts using lognormal distributions fitted to annual ratios between postseason abundance estimates and preseason forecasts, or constrained to assume unbiased forecasts. Accuracy was modest to low, with CVs exceeding 50% for 8/19 Chinook and 17/17 coho stocks. We evaluated the fitted median as a bias correction, and uncertainty buffers based on quantiles below the median. We tested whether retrospective application of bias corrections and/or buffers brought forecasts closer on average to postseason estimates; and performed retrospective and prospective analyses of consequences for stock status, harvest, and escapement for Sacramento River Fall Chinook (SRFC), a key fishery stock. Bias corrections and/or buffers improved most forecasts, with buffers providing improvement more often. For SRFC, bias correction alone could have led to one less year of overfished status, while buffers could have further shortened or avoided overfished status and reduced the frequency of under-escapement. Reductions in mean annual harvest resulting from applying bias corrections and/or moderate buffers were predicted to be smaller than the increases in harvest resulting from forecast and implementation error. Prospective simulations showed buffers could reduce risks of overfished status and under-escapement, at small costs to long-term mean harvests. However, this metric misses substantial harvest reductions in some years, since mean harvest is most sensitive to harvest at high abundance; though our analyses also neglected benefits of increased escapement for future production. Future work should incorporate observation error and nonstationarity, and the combined effects of forecast and implementation error on the probability of missing escapement goals.