In general, large-scale deterministic components, such as
trends and seasonal variations, should be eliminated from
the inputs since the network will attempt to learn the trend
and use it in the prediction (
Nelson et al., 1999
;
Pantazopoulos et al., 1998
).