Forecasting Trends
Predictive skill varied among the three models across the forecasting
period. The majority of the data (~97%) during this
period consisted of logbook records (Figure 3), reflecting the
seasonality of the albacore fishery in the California Current, which
operates primarily between May – November (Xu et al. 2017). Among the
models, predictive skill varied primarily based on spatial dependence
treatments (i.e., spatially implicit or explicit). GF and iSDM exhibited
similar trends, performing best on average in 2014 and 2019, with their
poorest performance in 2016 (Figure 3a,b). HE followed similar patterns,
but consistently showed lower performance, with its poorest performance
in the beginning of the forecasting period (Figure 3a,b). Furthermore,
the degree of novelty for each variable was correlated with the patterns
of AUC (r = -0.62 and -0.58 for MLD and SST, respectively) and
MAE (r = 0.7 and 0.55 for MLD and SST, respectively), equating to
poorer predictive skill as conditions became increasingly novel relative
to the training period (Figure 3c). For example, months experiencing the
highest degree of novelty in MLD and/or SST (e.g., May 2018, April 2016,
and January 2016) exhibited the poorest model performance, particularly
for the HE model (Figure 3). Trends in novelty for MLD and SST were
similar across all years, with small differences in Hellinger Distance
between MLD and SST (Figure 3c), suggesting both variables exhibited
comparable levels of novelty and variability during the forecasting
period relative to the training period.