In my thesis work I was making research in demand planning. This process plays key role in the area of manufacturing industries. The main purpose is to keep demand and supply in balance. In my thesis the concept of Supply Chain Management (SCM) is presented. I analyze the widely used methods and techniques for demand forecasting. I study the theory and application possibilities of neural networks, and I use them for the given task. The main goal of my thesis is to create a neural network for demand forecast and to compare its performance to classical forecast methods. For comparison among many others I used the well-known ARIMA method as baseline model.
Deep neural networks aren’t commonly used in demand planning yet. In my work I used novel deep learning methods. For implementation I used the Keras framework and Python programming language. First I determined and explored the necessary data sources, explored the data and I’ve done feature engineering. During my work I applied different kinds of neural network architectures and evaluated them. Furthermore, I tried other structural variations, for example introducing embedding layer for category features. I was working with two different databases. From one hand it was public sales data of a German drug store, Rossmann, provided by Kaggle.com. On the other hand I used private data of one of the SAP customers (provided by SAP). Both of these datasets contain historical data of customer actions. The difference that the first one is about sales data, the second is about customer orders. It was challenging to find the ideal hyperparameters of the neural net, and to do feature engineering as well.
In my thesis work I created deep neural networks for demand forecast that can handle several product types in one. For better accuracy, I also processed online posts from social network and created an ensemble model for sentiment analysis. The results and the performance of the different models were evaluated and compared to each other.