The forecasting methods I am interested in include: croston, naive, seasonal naive, ma, adida, exponential smoothing, TSB intermittent demand method).
In case the time series contains a seasonal component or an overall trend, is there a rule of thumb about which method to use? To my understanding, one should always deseason the data and then apply forecast methods. Is that right or should some methods be applied to the original time series (eg seasonal naive, exponential smoothing)?
Thank you!
Deseasoning, then modeling and forecasting, then reseasoning is one possible approach. It is by no means "always" done. For instance, you may run into problems if your seasonality interacts with other dynamics. An example is provided by the Air Passengers data set, where the seasonal oscillations increase over time, because they interact multiplicatively with the trend.
R code: plot(AirPassengers)
Of course, you could also have interactions between seasonality and other drivers.
There are many more sophisticated approaches, like forecast::ets() for R, which will attempt to automatically find an appropriate exponential smoothing model. For the above time series, it will fit a model with multiplicative error, additive dampened trend and multiplicative seasonality.
I recommend Forecasting: Principles and Practice (2nd ed.) by Athanasopoulos & Hyndman and Forecasting: Principles and Practice (3rd ed.) by Athanasopoulos & Hyndman as a textbook.
