Forecasting the monthly abundance of anchovies in the South Sea of Korea using a univariate approach

Abstract

Japanese anchovies, Engraulis japonicus, have exhibited substantial fluctuations in production in Korean waters. Anchovy drag net and drift gillnet fisheries are the major types of fishery that target different sizes of anchovies during different fishing seasons. We analyzed the monthly catch per unit effort (CPUE) for 1987–2012 using exponential smoothing methods, seasonal autoregressive integrated moving average (SARIMA), autoregressive neural network (ANN), and autoregressive recurrent neural network (ARNN) models to forecast the abundance of anchovies based on the fishing conditions. For the drag net fisheries, SARIMA provided better statistical insight than the other models, but the ARNN model was best for future forecasting (r2=0.819, PI=0.733). Additive Winters’ exponential smoothing (AWES) was the most effective of the three smoothing methods, but its validation was poor compared with SARIMA and the other neural network models. AWES, SARIMA, ANN and ARNN were less suitable for the drift gillnet fisheries. A comparison of the CPUE of drift gillnet fisheries to that of drag net fisheries showed that the data had been contaminated by factors such as periods when anchovy fishing was prohibited or a fishery had been abandoned due to economic reasons, rather than biological factors. Nevertheless, ARNN proved to be an effective and accurate model in the training phase, and its forecasts showed a comparatively strong statistical performance (r2=0.797, PI=0.662) in the context of short-to-medium-length time periods. Additionally, whereas SARIMA performed worse than ANN or ARNN, its forecasting capability was comparatively satisfactory (r2=0.713, PI=0.584). Finally, SARIMA has the advantage of providing statistical descriptions of the catches.