0In recent past when time series forecasting is done using Machine/Deep Learning techniques, forecasters are either not understanding or paying minimal attention towards imbalanced time series (unusual difference between normal and special time period values) which is regular phenomenon; for instance, electricity consumption, traffic predictions, footfall predictions to stores, etc. Machine/Deep Learning models often struggle with these imbalances because they lean heavily on the training data predominant patterns. Holidays and special events, being infrequent, don't leave a significant footprint in these models. Hence, generation of temporal features and considering time-varying models provide more accurate forecasts and also marketing effectiveness.
1) First understand both uni-variate and multi-variate context.
2) Second, what matters lot from domain expertise before modelling.
3) Feature Engineering and Selection
4) ML /TS Techniques
5) Without Validation Forecasting is incomplete.
Write up and visualizations.
