As far as I understand, it is always a good idea to apply a smoothing technique to a raw time series before training a model with it. However, I have a time series with big changes in magnitude and strong seasonality.
When applying a smoothing algorithm, such as MA or EWMA, the resulting time series does eliminate part of the noise, but also it stops correctly capturing these big changes in magnitude that are not noise.
Is there a smoothing algorithm that eliminates the small white noise, but does not interfere (as far as possible) with the shape of the time series? Furthermore, is it actually necessary to apply some smoothing technique to the raw data before training the model?