Prediction of consumption based of customers' segmentation

点击翻译按钮获取中文摘要

Strategic forecasts are best developed by combining historical data, collaborative supply chain information, and qualitative inputs. However, supply chain partners are not always willing or able to share information, and access to knowledgeable individuals for qualitative inputs is sometimes difficult since they are often senior management. This limits the information sources to historical data, processed through time-series analysis. While there have been many advances in the field of time-series forecasting over the past 40 years, the techniques all share the same limitation that they attempt to use past events to predict the future; hence, time-series analysis can only generate accurate forecasts for a few future periods. The best way to understand customers' behavior would be to monitor each customer ... even if this technique is applied in some specific domains (Hydro Quebec with its smart meters, for example), it is not always possible. Then each company needs to "evaluate" the customer behavior based on available data. Some data that are often available are delivery data. Delivery data may contain the product description, quantity delivered, customer address, and sometimes other information.The focus of this project is to develop methods and tools to analyze delivery data (what, how many/much, where) in order to get a better understanding of the customer, and propose a highly efficient customers' requirements prediction tool.Different questions need to be answered: Q1) how to segment customers with similar behaviors, Q2) how to transform delivery data to usage data, Q3) how to predict requirements for a segment, introducing external data, and Q4) how to derive predictions for each customer. None of those questions are actually completely solved. The scientific community already provides partial answers to each of them, but available tools need to be improved and integrated all together in order to provide performant answers.This project is developed in collaboration with an industrial partner that delivers bulk materials.