180 
Fishery Bulletin 109(2) 
cies (Lloret et al., 2000). The data for meagre are 
autocorrelated and present a relatively stable seasonal 
pattern. Also, the meagre is long-lived and a targeted 
fish in central Portugal (Prista et ah, 2009; Prista et 
al. 2 ). Therefore, it is possible such features contributed 
to the good forecasts obtained from the SARIMA model. 
However, we note that the landings of many short-lived 
pelagic species and species with variable seasonal pat- 
terns have also been well forecasted with SARIMA 
models (Stergiou, 1990a; Stergiou et al., 1997; Geor- 
gakarakos et al., 2006; Tsitsika et al., 2007) and that 
the meagre landings also display substantial annual 
and monthly stochasticity Therefore, such general pat- 
terns should not be considered as strict limitations to 
SARIMA forecasting. More importantly, we note that 
SARIMA models can forecast well only if they have 
been adequately identified and estimated, and always 
under the assumption that the future is behaving like 
the past (Chatfield, 1993). Consequently, factors like 
data quality, presence of outliers, and model selection 
criteria are also very important for model performance. 
We discuss these next. 
The quality of the input data for SARIMA models 
is determined mainly by the temporal stability of the 
statistical properties of the fisheries process and the 
consistency of its sampling over time. Consequently, 
although accuracy is required for some model appli- 
cations (e.g., Zhou, 2003), data inaccuracies do not 
necessarily undermine SARIMA forecasts as long as 
factors such as fishing practices, regulatory measures, 
or data collection practices can be assumed to remain 
constant. When dealing with shorter series, a care- 
ful check whether these assumptions hold becomes 
particularly important because model identification 
and estimation are very dependent on the few obser- 
vations available (Hyndman and Kostenko, 2007) and 
statistical techniques used to incorporate the effects 
of process changes in the models (e.g., Fogarty and 
Miller, 2004) are difficult to implement. In the case 
of meagre, the use of a short and recent time series 
better supported the assumption that data collection 
procedures, fishing techniques, fishery regulations, 
unreported landings, discards, and law enforcement 
practices did not change over time. In contrast, it is 
probable that these assumptions were not met in some 
less successful applications of the model to longer time 
series (e.g., Park, 1998). 
Outliers are known to cause trouble in time series 
model identification, estimation, and forecasts — an ef- 
fect that is amplified in shorter time series (Chatfield, 
1993; Trfvez and Nievas, 1998). The effects of outliers 
on forecasting performance are most disastrous when 
they occur near the forecasting origin because there 
they not only condition model structure and parameter 
estimates but are directly incorporated into the fore- 
casts (Chatfield, 1993). The meagre data set presented 
no apparent outliers and this likely contributed to the 
good fit and forecasting performance achieved. If outli- 
ers were present, specific modeling techniques could 
have been used to estimate their influence, smooth 
them, or incorporate them into the model (e.g., Chen 
and Liu, 1993; Lloret et al., 2000). We note, however, 
that any outlier during the hold-out period could still 
have changed our perception of model performance, 
even if it did not compromise the overall adequacy of 
the SARIMA model to forecast the landings. 
In time series analysis, adequate model specification 
is considered the most important driver of forecasting 
accuracy (Chatfield, 1996b). The difficulties of specify- 
ing an appropriate model increase for data sets with 
lower information content, such as those of highly vari- 
able short time series from more complex processes 
(Hyndman and Kostenko, 2007; Appendix 2). To date, 
fisheries applications of SARIMA models have essen- 
tially relied on Box-Jenkins (BJ) model selection pro- 
cedures to specify a model, and models with p <2 and 
q <2 have generally been selected (e.g., Mendelssohn, 
1981; Pajuelo and Lorenzo, 1995; Lloret et al., 2000). 
Compared to these, the model for meagre seems over- 
parameterized, but we note that all of its parameters 
are statistically significant and that the low RMSE /i)rec 
to RMSE /(f ratio indicates an excellent correspondence 
between fit and forecasting performances (Chatfield, 
1996b). In fact, although reduced model parameteriza- 
tion is considered beneficial to accuracy in forecast- 
ing, the most important aspect of time series analysis 
is not the number of parameters, but the degree to 
which the model approximates the statistical process 
underlying the data and whether or not it achieves 
the forecasting objectives (Chatfield, 1996b; Burnham 
and Anderson, 2002). In the case of meagre, had Box- 
Jenkins procedures been used, the selected models 
would be simpler and would still adequately fit the 
data: (l,0,0)x(l,l,0) 12 or (0,0,l)x(0,l,l) 12 . However, they 
would have performed worse than our AIC c -selected 
model in most performance metrics (RMSE: 0.245 and 
0.302, APE: 1.7-92.7% and 20.6-72.4%, MAPE: 44.1% 
and 44.0%, PE: 13.7% and 31.7%, respectively). These 
results show the impact that different model selec- 
tion techniques may have on forecasting performance 
with SARIMA models and stress the importance of 
considering objective data-driven criteria like AIC f for 
circumventing the subjectivities of model selection in 
smaller data sets (Hurvich and Tsai, 1989; Burnham 
and Anderson, 2002). 
Conclusions 
Use of SARIMA models in monitoring fisheries 
From a strictly forecasting perspective, SARIMA models 
have often been criticized for the excessive reliance on 
past time series behavior and their difficulty in predict- 
ing future structural changes (Georgakarakos et al., 
2002; Koutroumanidis et al., 2006). Our results show 
that these drawbacks can become major advantages 
when SARIMA models are used for monitoring fisher- 
ies. At present, none of the European meagre fisheries 
is subjected to routine analytical assessment. By fitting 
