Time series analysis for business analytics

In the last years we have seen an explosion of research and innovation in the area of Business Analytics (BA), and a comparable trend is verifiable in the industry in terms of solutions and services splashing from all remote parts of the world, all of them raising the ’analytics’ flag. Apart from the fact of the generalized abuse of this buzz word, some underlying prompts have to be uncovered and properly addressed: • Companies have to cope with constant change and uncer- tainty, more than ever before. • InformationandCommunicationTechnologies(ICT)andthe Internet are the key drivers of the globalized world we have nowadays. • The amounts of data produced and accessed by companies and individuals largely exceed the capability of a proper value-extraction processing in the vast majority of the scenarios. Advancements in ICT and the Internet bring us tons of data with almost no charge. However, we do lack the right procedures to digest these data oceans, and even more importantly, to extract the value hidden in them and timely contextualize it in the right scenarios. This general consideration puts on the table the main components of the problem BA attempts to solve. Firstly, we refer to a business layer, full of semantic and meaning, which includes the abstract concepts taking part in business administration and operation. Executives and managers think in terms of sales, revenue, costs, budgets and projections. Business strategy is defined in terms of plans, goal maps, balanced scorecards and business operations. In the other hand we find the counterpart, which corresponds to the data layer. ICT revolution makes it possible to track and store tones of data about daily operations in every business. This digital universe exponentially grows with time. But although it is widely known that information is one of the most valuable assets of today’s organizations, bridging the abstraction gap between data lakes and business concepts remains the main obstacle for a more widespread implementation and exploitation of BA in most industries and sectors. Closing the conceptual gap between raw data and the business concepts is a very challenging problem that captured attention from the research community since the first emergence of BA systems. This thesis contributes to this aim in several ways. Firstly, a new toolbox for time series analysis based on state space modeling has been developed and published under the name SSpace. SSpace is a MATLAB toolbox that provides a number of routines designed for a general analysis of SS systems. It combines both flexibility and simplicity, and at the same time it enhances the power and versatility of SS modeling in a friendly environment. The toolbox possesses very distinct properties to other SS pieces of software, but at the same time takes advantage of methods and algorithms from other sources, mainly Taylor et al. (2007) and Durbin and Koopman (2012) as it has been unfolded in Chapter 3. The combination of all these factors gives SSpace a particular flavour. This toolbox serves as the basis for all the rest of developments in this thesis. Secondly, new methods for time series automatic identification has been developed in order to tackle with the problem of analysing time series in a BA context, i.e., companies with an accumulation of so many time series to model and forecast that detailed manual identification becomes essentially impossible (Fildes and Petropoulos, 2015). The short life-cycle of the analysis and the acceleration of business operations contribute still more to the need for automatic identification techniques. The contribution of this thesis to the field of automatic identification is twofold. On the one hand, a general approach for model selection is proposed that combines different criteria with additional information of the time series itself as well as the responses and fitted parameters of the alternative models. On the other hand, a much more concrete contribution consists of proposing a novel automatic identification procedure of a particular class of models, namely the Unobserved Components (UC), as a valid candidate to compete with other classical alternatives in the field of forecasting time series in BA contexts. It is interesting to note that the UCs have been systematically disregarded in a big part of the forecasting literature because of the overwhelming presence of ETS and ARIMA methods, while their potential is immense, due to thei