I trained the LSTM model to forecast the number of people infected with COVID-19. Since each country and state (geo location) has different number of population as well as the infected over time, I normalised each time series data by/within geo location (using min-max scaler). So, the model can learn change in time series in the same scale. After the normalisation, my model appears to have lower error rate than when it was NOT normalised by/within geo location.
However, I wonder if this is a fair way to create universal forecaster to predict for multiple countries. I saw some does forecasting after time series clustering instead. Do you think that is better approach?
Here is the link for my solution.
Please share your thoughts on this. Thanks!