MODELING AND FORECASTING ANNUAL FISHERIES CATCHES - COMPARISON OF REGRESSION, UNIVARIATE AND MULTIVARIATE TIME-SERIES METHODS

In the present work, eight forecasting techniques are evaluated on the basis of their efficiency to model and provide accurate operational f orecasts of the annual commercial landings of 16 species or groups of species in the Hellenic (Greek) marine waters. The development of oper ational forecasts was based on the following four general categories o f forecasting techniques: (a) deterministic simple or multiple regress ion models incorporating different exogenous variables (time-varying r egression, TV; multiple regression models incorporating time, fishing effort, wholesale value of catch and climatic variables, MREG); (b) un ivariate time series models (Brown's one parameter exponential smoothi ng, BES; Holt's two parameter exponential smoothing, HES; and AutoRegr essive Integrated Moving Average (ARIMA)); (c) multivariate time serie s techniques (harmonic regression, HREG; dynamic regression, DREG; and vector autoregressions. VAR); and (d) the 'biological' exponential su rplus-yield model, FOX. Fits (for 1964-1987) and forecasts (for 1988-1 989) obtained by the different models were compared with each other an d with those of a naive method (NM) and an empirical one (i.e. combina tion of forecasts, EMP) using 32 different measures of accuracy. The r esults revealed that HREG and MREG models outperformed, in terms of fi tting accuracy, the remaining eight models (NM, TV, BES, HES, FOX, ARI MA, VAR and EMP). They were both characterised by: ( a) higher accurac y in terms of all, or most, standard and relative statistical measures ; (b) unbiased fits; (c) much better performance than NM; (d) transfor med errors which were essentially white noise. In addition, HREG and M REG models: (e) explained from 79% to 97% of the variability of 13 tra nsformed annual catches and from 31% to 61% for the remaining ones; (f ) produced fits with MAPE values ranging from 3.4% to 21.2%; (g) in al l, or most cases, predicted the between year variations during the fit ting period, 1964-1987. In terms of forecasting performance, however, not a single best approach was found for the 16 annual catches. In gen eral, BES and, to a lesser extent, HES, NM ( which actually is an ARIM A (0,1,0)), EMP and HREG models were among the best performers more of ten, produced the worst forecasts more rarely and were generally chara cterised by the higher number of stable forecasts and of forecasts wit h MAPE < 20% and < 10%. TV was also efficient for some annual series. Conversely, the poorest performers (FOX, MREG and ARIMA) rarely did be tter than average. The biological FOX models produced the least accura te and biased fits, bad forecasts (> 34.1%) in two out of four cases, and were characterised by transformed errors that were significantly ( P < 0.05) autocorrelated. Some of the empirical models also had intere sting explanations. Hence, the univariate ARIMA and multivariate VAR t ime series models predicted persistence of catches. The multivariate V AR and HREG time series models also predicted cycles in the variabilit y of the catches with periods of 2-3 years. Moreover, FOX and MREG mod els indicated that fishing effort, wholesale value of catch and climat e may, in a synergetic fashion, affect long-term trends and short-term variation in the catches of, at least, some species (or groups of spe cies). Finally, MREG and VAR models predicted that variability and rep lacement of anchovy by sardine catches are not due to chance, and wind activity over the northern Aegean Sea may act as a forcing function.