A matter of time - leveraging time series data for business applications. : A matter of time - leveraging time series data for business applications

The global topic of this dissertation is nicely reflected by its title - 'a matter of time' -, as we explore the different uses of data, i.e. 'matter', that is collected sequentially in time and how this data type can be integrated and leveraged within business contexts. Time series data represents the changes that a given feature goes through over time and can be collected for any target or explanatory variable. Time series have therefore been part of business applications for decades and have been studied extensively in both academic literature and practical use cases. This work includes both sides of the spectrum and tries to combine both theoretical and practical points of view on the subject in each project. This thesis therefore has two focal points: business-oriented research, which aims to integrate business concepts into time series use cases in general, and application-based research, which investigates how time series data can be leveraged in specific case studies. Within these two sections, we define three main research topics: profit-driven forecasting, model automation and benefiting from time series data in applications. In terms of profit-driven forecasting, we introduce an expected profit function that can be used for sales forecasting applications, which was extensively tested on datasets from The Coca-Cola Company. If forecasting models are not able to produce accurate forecasts, this leads to an increase in costs, e.g. overstock costs. We therefore calculate the expected profit of a forecast by combining forecasting accuracy with a profit margin that can be set by a business expert in order to evaluate a model in an entirely different way. The resulting formula is easily adaptable to any situation and interpretable as the potential monetary value of a forecast. We show that this function can be integrated into both the building and evaluation of a forecasting model, and can be used next to traditional evaluation metrics for the final model selection. The expected profit function therefore emphasizes the value of a forecasting model by connecting a well-known practical concept, i.e. profit, with the evaluation of existing techniques, which also facilitates its integration into forecasting processes that are already in place. The second topic of the business-oriented research looks into the automation of forecasting models specifically and analytics in general. Model automation can be useful for two separate reasons in business contexts. First of all, if there are too many analytical models to train, it could be beneficial to find a data-driven way of parametrising these models based on their input. By automating this process, less analytics experts are needed and the models could be shared more easily across different people or departments. In this thesis, we therefore investigate how to combine the expected profit function with automatically selecting parameters and features for sales forecasting applications. We even try to optimize the search for the correct parameters instead of using the brute force approach by turning to evolutionary algorithms. The second use for automated analytics is especially apparent in cases where there is a complete lack of analytical expertise, e.g. in smaller companies. Analytics-as-a-Service (AaaS) offers a potential solution to this problem by offering web-based platforms that could also be used by business users to implement ready-made predictive models. Our research takes a closer look at the potential users of these AaaS tools and concludes that some novice users are able to achieve sufficient results if they meet certain requirements. The final topic of this dissertation is difficult to generalize, as all of its conclusions are inevitably linked to the application of the work in question. The two projects of this section deal with completely different use cases: blood donation forecasting and churn prediction by time series classification. In the first study, we leverage the hierarchical structure of the time series data to generate better forecasts by using a bottom-up approach. Furthermore, we look into the potential of web traffic data to capture the intention of a donor to donate in the near future. In the second project, we turn to time series as a means to deal with dynamic features for a churn prediction application. In a world where a large amount of data is being created and collected every day, we have to become more aware of the dynamic aspect of all of this data. For example, a lot of analytical tasks take user behaviour into account and assume that the most recent data about a person has the highest predictive value. However, we wanted to examine whether capturing a change in this behaviour does not lead to a better performing model in the end. Concretely, we take a closer look at a classification use case that aims to identify churners for a telecommunication company by looking at changes in their call detail records. In short, this dissertation looks into the different uses of time series data in several business applications. This work shows the possibilities for this type of data, which is already present in a lot of businesses, but has not reached its full potential yet. status: published