Hasso-Plattner-Institut
Prof. Dr. h.c. mult. Hasso Plattner
 

Predictive Analytics on In-Memory Databases

Motivation

For manufacturers it is important to have an accurate demand forecast for their products in order to avoid over or under capacity in their stores. In case of Vendor-Managed-Inventory the manufacturer is solely responsible for filling the shelfs inside the retail stores. Point-of-Sale (POS) data is one of the most important basis for forecasting. However, for different reasons, many shops cannot provide this kind of data. Instead of using imprecise shipment forecasting, new approaches have to be evaluated. 

The in-memory technology, which has been developed for more than four years at our chair, allows the efficient analysis of big-data. In detail, the technology provides compression rates of data that allow to keep even large data sets inside the main memory. Furthermore, in-memory leads to high performance analysis possibilities, at the same time creating an unprecedented level of flexibility. These features afford completely new interaction scenarios with transactional systems that lead to a total rethinking concerning possibilities and chances.

Project Description

 The goal of this project is to get an overview of existing forecasting methods in the context of massive data sets as well as adapting and optimizing these processes for modern in-memory technology. Within this context we evaluate a new approach based on a statistical model provided by the Massachusetts Institut of Technology. It allows to determine the relationship between POS and shipment data by using stores with similiar characteristics. Based on this relationship missing Point-of-Sale can be generated.

The first part of the project serves as an orientation phase in which the students acquire the understanding of the main concepts and ideas behind the in-memory technology. In addition to that an exploration and understanding of the core concepts of different smoothing and prediction algorithms is intended. Therefore, it is necessary to analyze data to identify structures and models, which are accurate for applying these algorithms. 

The second part of the project enfolds the practical implementation of the gathered knowledge. The result of this process is a web application which provides a user interface allowing the complete workflow of a demand forecast within the theoretical border of the first phase. In order to benefit from the priorly analzyed concepts of in-memory technology, the team members construct their prototype on SAP HANA, which allows to handle vast amounts of data within accurate time. The application should exemplify in how in-memory technology can help optimizing predictions in order to improve existing supply chain processes.