CORDIS - EU research results

Big Time Series Analytics for Complex Economic Decisions

Periodic Reporting for period 1 - BigTime (Big Time Series Analytics for Complex Economic Decisions)

Reporting period: 2019-05-01 to 2021-04-30

Large, complex datasets (‘Big Data’) are nowadays omnipresent in business and economics. Their variety and sheer size provide nearly endless opportunities to improve economic decision making at European governments, companies and universities: social media and internet search data could shed light on consumer sentiment, payment transactions could be informative for private consumption, or road traffic data might reflect trade developments. The aim of this action is to make new statistical methods available for more confident economic decision making through statistical learning from big data. Special emphasis is placed on developing methods for analyzing big time series data as many real-life problems have a time component.

While the Big Data concept has been around for years, most organizations now understand they can get significant value through better and/or faster decision making from Big Data. To this end, it is essential that the European Research Area has the relevant expertise from its statistical community to deliver new Big Data technology. Over the recent years, there has been a surge in novel statistical learning methods to analyze large data sets. Yet, analyzing large economic data often pose additional challenges in that many real-life economic problems have a time component that needs to be appropriately addressed.

This action treats time dependence not as a nuisance but rather a valuable source of additional information that is leveraged in the statistical analysis. The action’s main objective consists of tackling econometric challenges of developing statistical learning-based Big Time Series Analytics for economic data. It builds a partnership between statistics, machine learning and econometrics. In doing so, it contributes to the society’s ability to enhance economic discovery from Big Time Series Data by equipping researchers, business analysts, policy holders and students with a Big Time Series software toolbox.
The action’s overall goal is to develop new methods for accurate time series analysis of big and complex economic problems. Three specific Research Objectives have been tackled:

1. Honest methods for inference. The action delivers accurate and reliable measures of uncertainty quantification for high-dimensional time series problems. First, bootstrap unit root tests are provided (Smeekes and Wilms, 2020). Unit root test from an essential part of any time series analysis. The software package bootUR provides practitioners with a single, unified framework for comprehensive and reliable unit root testing on single time series or potentially a large number of time series (including panels). Secondly, inferential procedures for large time series models via the desparsified lasso are developed (Adamek, Smeekes and Wilms, 2020).

2. Estimation of Impulse Response Functions. The action delivers novel tools to describe how the economy responds, over time, to unpredictable events (called shocks). As such, policyholders and other decision makers can anticipate outcomes that have not yet occurred. Such tools crucially hinge on accurate estimation procedures for high-dimensional time series models, which have been developed in Barbaglia, Croux, Wilms (2020), Hecq, Ternes, Wilms (2021), Nicholson, Wilms, Bien, Matteson (2020) and Wilms, Rombouts, Croux (2021) and are made publicly available through the software package bigtime. Further extensions towards impulse response analysis based on, for instance, local projections are work-in-progress.

3. Identification strategies for high-dimensional time series models. The action delivers identification strategies that map observed economic data to the relevant economic parameters of interest. Such strategies should be applicable in high-dimensions, favor parsimonious models and integrate identification and estimation strategies via regularization procedures. Wilms, Basu, Bien and Matteson (2020) deliver such strategies for vector autoregressive moving average models; other time series models such as structural vector autoregressive models are currently being studied.

Concerning exploitation and dissemination, several activities have taken place. The work performed during this action has been presented at seminars (Maastricht University 2019; University of York, 2019; Erasmus University Rotterdam, 2020); conferences (Joint Statistical Meetings 2019; CMStatistics 2019; NESG 2020; (EC)^2 2020) and workshops (Big Data and Forecasting Workshop at the Joint Research Centre 2019; StatScale Workshop 2021). The Online Workshop on Dimensionality Reduction and Inference in High-Dimensional Time Series marks the end of the action. Besides, software toolboxes have been made publicly available on CRAN via the R packages bootUR and bigtime.

Adamek R., Smeekes S. and Wilms I. (2020), Lasso inference for high-dimensional time series, arXiv:2007.10952.
Barbaglia L., Croux C. and Wilms I. (2020), Volatility spillovers in commodity markets: A large t-vector autoregressive approach, Energy Economics, 85, UNSP 104555.
Hecq A., Ternes, M. and Wilms I. (2021), Hierarcical regularizers for mixed-frequency vector autoregressions, arXiv:2102.11780.
Nicholson W.B. Wilms I., Bien J. and Matteson D.S. (2020), High-dimensional forecasting via interepretable vector autoregression, Journal of Machine Learning Research, 21(166), 1-52.
Smeekes S. and Wilms I. (2020), bootUR: An R package for bootstrap unit root tests, arXiv:2007.12249.
Wilms I., Basu S., Bien, J. and Matteson D.S. (2020), Sparse identification and estimation of high-dimensional vector autoregressive moving averages, arXiv:1707.09208.
Wilms I., Rombouts, J. and Croux C. (2021), Multivariate volatility forecasts for stock market indices, International Journal of Forecasting, 37(2), 484-499.
This action has the potential to create a widespread impact as Big Time Series problems arise across many disciplines. By illustrating the usefulness and need for the developed methods on a wide variety of application domains in macro-economics, finance and marketing, direct spill-over effects towards these scientific areas have been assured.

The recent financial and euro area sovereign debt crisis have boosted efforts at, for instance, central banks to obtain richer “Big Data” sets for policy analysis. In his “Policy Analysis with Big Data” speech, Benoît Coeuré (Member of ECB’s Executive Board), indicated that the push towards more fine-grained data poses various econometrics challenges. The project directly tackles (some of) these challenges by developing methods that allow policy makers to reveal estimates of policy and/or private agents’ behaviour and their impact on the economy. The project is thus of considerable importance to financial/economic institutions such as central banks and the outcome is likely to create societal impact on governments and policy makers through them.

Besides, the action also considers industry partners as important stakeholders. Maastricht University’s Institute of Data Science and the research group on Data-Driven-Decision-Making facilitated networking activities among academia, business and industry. To further enhance the project’s impact and visibility to a broader audience, an online course on Time Series Analysis has been delivered which specifically aims at informing industry practitioners on the latest state-of-the-art research methodologies via a hands-on approach.
Network Analysis of Large Dynamic Systems