748 



Fishery Bulletin 94(4). 1996 



released to the industry) is included as well (Table 

 2) because Hurricane Andrew impacted south Florida 

 in August 1992 and disrupted ENP data collection 

 through the remainder of 1992 and into 1993. All 

 data were held until ENP had verified that record- 

 ing instruments had not been disturbed by the hur- 

 ricane or that instruments had been resurveyed and 

 appropriate corrections had been applied to data 

 sets. 7 These revised forecasts are presented in Table 

 2 because ENP data figured prominently in every 

 other model. 



All models derived in 1987, 1990, 1991, 1992, and 

 1994 produced forecasts that were within ±20% of 

 actual landings (Table 2). All 1990 and 1991 models 

 (models 7-12) contained the same first five variables. 

 Models 13 and 14 (1992) and models 22 and 23 ( 1994) 

 contained similar independent variables as did suc- 

 cessful 1990-91 models. Forecast models for 1989 

 included two relatively successful ones (models 3 and 

 4, -10% and +18% of actual landings) and two poor 

 ones (models 5 and 6, +45% and +44% of actual land- 

 ings). None of the 1989 models approached the 

 Mallow's test criterion of C = p, but models 3 and 4 

 had relatively high R 2 values compared with those 

 of models 5 and 6. Model 2 (1988) and models 15-17 

 (1993) produced unsuccessful forecasts. Model 2 was 

 the only forecast equation that had no variable di- 

 rectly associated with the Tortugas fishery, although 



all other unsuccessful models did have such vari- 

 ables. Models 15—17 were derived without access to 

 ENP data which had been incorporated in all other 

 models, successful or otherwise. Had ENP data been 

 available, models 18—21 would have been generated 

 with all but model 19 producing successful forecasts. 

 Models generally forecasted increases or decreases 

 in future landings correctly, with exceptions in 1989, 

 1991, and 1993. In both 1989 and 1993, three of four 

 models correctly forecasted increased landings, 

 whereas in 1991 only one of three models correctly 

 forecasted decreased landings. Finally, the forecast 

 models for 1995 (for which actual landings will not 

 be available until late 1996) are presented for com- 

 parison with previous models and as documentation 

 of the wide range in forecasts (2,858-4,581 t) seen 

 previously only in 1989. 



Every model contained at least one independent 

 variable measured in September, and some models 

 included October measurements. These data were 

 usually unavailable until November, thus my goal of 

 releasing each forecast in October was never achieved. 



The fact that several variables consistently entered 

 forecast models that were generated annually argues 

 for selecting a single model with fixed variables. This 

 strategy was assessed by applying monthly data, 

 collected in years after generating each model, to 

 those models in order to determine accuracy, if a given 



