Machine Learning Approaches for Disaggregated and Intermittent Demand Forecasting for Last-mile Logistics

Demand forecasting is one of the most critical steps in planning. The success of the operations and customers' services highly depend on reliable forecasts; however, forecasting is not easy, especially when the time series is highly volatile. Intermittent and lumpy demands are even more challenging to forecast due to zero values in the observations. In last-mile logistics, when the product variety is high, both fluctuation and intermittency of the demand are faced. Forecasting such demand becomes extremely crucial. This paper uses machine learning techniques to forecast demand for 160 different products for 120 different clusters in Riyadh metropolitan area, Saudi Arabia. Our models forecast daily demand for each product in each region, predicting more than 4000 entries. Our results showed that the machine learning models perform much better than traditional methods like Croston's method.