Demand forecast models for online supermarkets
J.M. Evers (Student TU Delft, Picnic Supermarkets B.V.)
LA Tavasszy (TU Delft - Transport and Logistics, TU Delft - Transport and Planning)
Ron van Duin (Rotterdam University of Applied Sciences, TU Delft - Transport and Logistics)
DL Schott (TU Delft - Transport Engineering and Logistics)
Frank Gorte (Picnic Supermarkets B.V.)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Food waste and incomplete orders at online supermarkets mainly occur due to inaccurate demand forecasting which leads to incorrect ordering of products. The objective of this study is to develop an accurate demand forecast model at product-level based on historical customer order data, give recommendations on implementation and describe impact on logistical planning at an online supermarket in the Netherlands. The product of research in this case study was bread, because of the habitual order behaviour of customers.
It is found that, using historical customer behaviour, model accuracy can be increased by forecasting the bread order probability for every customer individually compared to a total order regression. Decision regression trees and random forest regression models are implemented to forecast product sales on short term and show high accuracy. The forecast accuracy of predicting the number of breads per day is about 99.9 percent, given the number of customers that is going to order. This implies that it is feasible to order bread directly at supplier with a significant level of reliability such that waste due to overprediction and incomplete orders due to underprediction can be reduced to acceptable levels.
The main advantages of using tree models for demand forecasting of products for an online supermarkets is the fast run time, accurate forecasts and easy interpretation. Within a few seconds, thousands of customers are analysed and conclusions can be drawn on the future demand of bread based on historical demand of each customer. Average breads per order in historical orders and number of orders are main predictors for future demand. Decision trees and random forest regression outperform linear regression in this case study.