[x] Train separate LSTM / fc models on all inputs, one model for each month.
[x] Check performance (mse / correlation): Are there months with higher / lower accuracy?
[x] Use ensemble of these models as predictor: Do we outperform former single LSTM / fc model trained on all inputs including months as additional input features?