Is it necessary to remove Seasonality while time series forecasting using ML methods ? Can't model learn it on itself?

Question

I think ML model can learn from seasonal variations also. But if we remove seasonal variations, model & add it back, then essentially, we will end up dividing learning into :

1. 'seasonal variations learning subtask'

2. ' learning relations of independent variables with target as another subtask'

Answer 1

There's no agreed upon answer to your question:

• Some sources recommend removing both the seasonality and trend, and letting the ML methods learn how to forecast the residuals effectively (See Smyl and the winning solution to the M4 competition for example).
• Others recommend removing the trend but keeping the seasonality.
• Others recommend throwing everything to the ML model, but maybe doing a power transform first to stabilize variances.
• Personally I have had good results by removing the trend and then normalizing using a MinMax scaler. The denormalizing the predictions and adding the trend component back.

Conclusion: Try all and see which one works best.