Feature Engineering to enhance Deep Learning forecasting of intermittent and lumpy demands: a comparative study

Forecasting intermittent demand using ML/DL tools faces challenges coming from the inaccurate detection of the future zero-demands resuting in low forecasting accuracy. The data attached represent the results of forecasting intermittent demands using three DL methods (GRU, LSTM, and TCN) while feeding these methods with different feature types. The aim is to detect the best feature scenario(s) for each method and the overall best forecasting methods. Besides, validating the positive impact of increasing the accuracy of zero-demand prediction on the overall forecasting performance.