Luke Mosley

PhD Project:

High-Dimensional Disaggregation of Time Series satisfying both Temporal and Contemporaneous Constraints

Partners:

STOR-i Centre for Doctoral Training at Lancaster University and the Office for National Statistics

Supervisors:

Professor Idris Eckley and Dr Alex Gibberd (Maths and Stats, Lancaster) and Duncan Elliott (Office for National Statistics)

Paper 1 (2021): Sparse Temporal Disaggregation

Abstract: Temporal disaggregation is a method commonly used in official statistics to enable high-frequency estimates of key economic indicators, such as GDP. Traditionally, such methods have relied on only a couple of high-frequency indicator series to produce estimates. However, the prevalence of large, and increasing, volumes of administrative and alternative data-sources motivates the need for such methods to be adapted for high-dimensional settings. In this article, we propose a novel sparse temporal disaggregation procedure and contrast this with the classical Chow-Lin method. We demonstrate the performance of our proposed method through simulation study, highlighting various advantages realised. We also explore its application to disaggregation of UK gross domestic product data, demonstrating the method’s ability to operate when the number of potential indicators is greater than the number of low-frequency observations.
ArXiv Link: https://arxiv.org/abs/2108.05783

Paper 2 (Current Work): Multivariate Temporal Disaggregation with Contemporaneous Constraints

Abstract: In economics, governments want to produce high-frequency information of regions in a nation to better understand regional disparity, deviations from the national avergae and regional correlation. This high-frequency information must agree both temporally with current low-frequency information and contemporaneously with national accounts. In hydrology, efforts are required to produce rainfall information at a more granular level, agreeing with the wider observed spatial coverage. In this work, we build a regularised multivariate regression model with a vector auto-regressive (VAR) error structure to produce high-frequency estimates of multiple time series that aggregtae to temporal and contemporaneous restrctions. The assumption of VAR errors allows one to model the correlation dynamics between regions in a data-based way. We incorporate regularisation techniques to estimate the large, sparse covariance matrix and allow for high-dimensional indicator sets to be used. We apply our method to annual UK NUTS 1 level regional GDP to produce quartery figures using a large indicator set. In addition, we perform a hydrology case study to disaggregate rainfall data.

R Package (Current Work): DisaggregateTS

We are in the process of building an easy-to-use R package that allows both univariate and multivariate disaggregation of time series. We make some of the well-established methods to do temporal disaggregation available, as well as the two methods we proposed in papers 1 and 2. These novel methods allow the user to include as many indicator series as they wish to perform the disaggregation task and an assessment of indicator relevance is returned. They also allow the user to perform a multivariate disaggregation of multiple time series that must satisfy contemporaneous constraints. Appealing graphical visualitions will be availale for the outputs of multiple time series. Real data sets will be available for the user to test out each method.

The MRes component of the STOR-i programme includes taught courses, projects and group activities providing me with a grounding in statistics and operational research, an overview of thriving research areas, and an opportunity to develop a formal research proposal for my PhD. Below is a list of project reports I have completed this year. Also, check out my blog to read about some of the research areas I have been exposed to.

16/04/2019: Using adaptive random search for simulation optimisation - POSTER

31/03/2019: Modelling the demand of healthcare systems using infinite-server queueing models

24/02/2019: Detecting multiple changes in variance of univariate time series data

03/02/2019: Continuous-Time Markov chains and Discrete-Time Markov decision processes

03/12/2018: Branch and Bound algorithm to solve the Knapsack problem