Reducing Uncertainty in Data-Poor Assessments Using Seasonal-Trend Decomposition

Full stock assessments are not possible for many data-poor stocks. Instead, so-called index-based methods are used in which fishery reference points are based solely on scientific surveys. Approximately 40% of stocks are assessed this way in the Northeast U.S. region. Although index-based assessments provide a convenient way of evaluating otherwise unassessed stocks, they are often plagued by high levels of uncertainty. Here we ask whether we can reduce this uncertainty using a well-established time series method known as STL decomposition. STL (an acronym for Seasonal and Trend decomposition using Loess) attempts to separate observation error from seasonality and long-term trend using an automatic Loess algorithm. Using fishery simulations and real data from over 100 stocks, we test the accuracy of STL-derived biomass estimates, reference points, confidence intervals, and forecasts. For both the simulations and real data we find that the STL method reduces the uncertainty and increases the accuracy of all assessment outputs.