Modelling and forecasting monthly fisheries catches: comparison of regression, univariate and multivariate time series methods

Abstract

In the present work, seven forecasting techniques were evaluated on the basis of their efficiency to model and provide accurate operational forecasts of the monthly commercial landings of 16 species (or groups of species) in the Hellenic marine waters. The development of operational forecasts was based on the following three general categories of forecasting techniques: (a) deterministic simple or multiple regression models incorporating different exogenous variables (seasonal time-varying regression, TVS; multiple regression models, MREG, incorporating time, number of fishers, wholesale value of catch and climatic variables); (b) univariate time series models (Winter's three parameter exponential smoothing, WES; ARIMA); and (c) multivariate time series techniques (harmonic regression, HREG; dynamic regression, DREG; vector autoregressions, VAR). Fits (for 1964–1987) and forecasts (for 1988–1989) obtained by the different models were compared with each other and with those of two naive methods (NM1 and NM12) and an empirical one (i.e. combination of forecasts, EMP) using 32 different measures of accuracy. The results revealed that the univariate ARIMA, closely followed by the multivariate DREG time series model, outperformed the others (NM1, NM12, TVS, MREG, HREG, EMP, VAR and WES) in terms of both fitting and forecasting accuracy. They were characterised by: (a) higher accuracy in terms of all, or most of the standard and relative statistical measures that were usually tied together; (b) unbiased fits and forecasts; (c) much better performance than NM1 and NM12. In addition, ARIMA and DREG models: (d) explained over 80% of the variance of the transformed catches; (e) had residuals that were essentially white noise; (f) in all cases predicted the amplitude and the start and end of the fishing season; and (g) produced forecasts that had mean absolute percentage error values under 28.2% for 11 out of 16 monthly series. The different measures employed also indicated that EMP and WES models outperformed NM1, NM12, TVS, MREG and HREG models. EMP produced forecasts with MAPE values under 23.2% for ten monthly series, whereas WES produced forecasts with MAPE values under 25.3% for eight monthly series. This suggests their potential use in short-term fisheries forecasting. The limitations of the different forecasting techniques, measures of accuracy and data used in the present study are also discussed. Some of the empirical models built also had interesting biological/oceanographic explanations. Hence, the univariate ARIMA and multivariate DREG and VAR time series models all predicted persistence of catches. The univariate ARIMA and multivariate HREG, DREG and VAR time series models all predicted cycles in the variability of the catches with periods of 1 and 2–3 years. Moreover, MREG, DREG and VAR models indicated that the number of fishers, wholesale value of catch and climate may, in a synergistic fashion, affect long-term trends and short-term variation in the catches of at least some species (or groups of species). Finally, DREG and VAR models predicted that variability and replacement of anchovy by sardine catches are not due to chance and wind activity over the northern Aegean Sea may act as a forcing function.