Which LSTM Training Strategy Performs better?
I would like to use LSTM for predicting multiple time series (Time series about sales per day in multiple countries. In parts, contradicting regional trends are present within the data. Sales is the 'dependent' variable, whereas there are multiple other explanatory variables). Now I'm thinking if I should train one one-fits-all LSTM model that uses the data of all countries for training or if I should train one LSTM Model for each country (there are 30 different countries in my dataset) separately.
My thoughts so far:
The 'one-fits-all' version can obviously incorporate more data points for training, as all countries are used for training. Thus, the model might be more robust in comparison to multiple separate LSTM models. As the sales data for each country not necessarily moves in the same direction, the multiple separate LSTM models solution might be more accurate (whereby probably facing the greater risk of overfitting). Any thoughts and insights (e.g., regarding accuracy, overfitting, ease of implementation)?
Thanks for your help in advance!
Topic lstm deep-learning neural-network python
Category Data Science