An increasing share of economic decisions is recorded as digital text, audio and video. At the same time, advances in computation and statistical inference have allowed for the exploitation of these unstructured data sources in scientific research. Such data can be used to extract useful information that is not available in traditional aggregate indicators of economic activity (e.g. stock prices inflation, or output). For example, text data from social media can provide incremental information that goes beyond traditional quantitative data (e.g. provide new measures political risk, political influence, economic sentiment), and they are timely, in that they are available at much higher frequencies than traditional economic data (daily instead of monthly/quarterly).
However, the challenges of modelling with text data are many. This research project focuses on two important contributions, relevant for prediction. First, the numerical representation of text data is inevitably ultra-high dimensional. While forecasting with large information sets is desirable, when more information translates into more parameters, this can be very hurtful (over-parametrization problem). Therefore, one dimension of the proposed research will examine statistical estimators (so-called “shrinkage estimators”) that prevent overparametrization, combined with computational algorithms that are of low complexity and are easy to use by practitioners. Second, the proposed research will focus on the interpretation of text data for prediction purposes. The current typical approach is to insert all keywords from a text document into a model that converts this big term-document matrix into a manageable indicator. However, such approaches are so-called “black-box” and little is known if, when constructing such an indicator, it will be relevant for the variable that we want to predict. Therefore, our intention is to examine procedures where indicators based on textual data are extracted in a way that there is always direct reference to the variable to be predicted.
Applicants must meet the following essential criteria:
• A good first degree (at least 2:1), preferably in economics, statistics, or computing science.
• Demonstrate an interest in, and knowledge of, natural language processing, high-dimensional estimation, and computational methods.
• Have a good grounding in economics, econometrics and finance.
Students must meet ESRC eligibility criteria. ESRC eligibility information can be found here*: https://esrc.ukri.org/skills-and-careers/doctoral-training/prospective-students/
The scholarship is available as a +3 (PhD only) or a 1+3 (MSc and PhD) programme depending on prior research training. This will be assessed as part of the recruitment process. The programme will commence in October 2020.
The award includes:
• An annual maintenance grant at the RCUK rate (in 2019/20 this is £15,009)
• Fees at the standard Home rate
• students can also draw on a pooled Research Training Support Grant, usually up to a maximum of £750 per year http://www.sgsss.ac.uk/studentship/economic-predictions/
Applications will be ranked by a selection panel and applicants will be notified if they have been shortlisted for interview by Monday 13 April 2020. Interviews will take place on Monday 27 April 2020.
All scholarship awards are subject to candidates successfully securing admission to a PhD programme within the University of Glasgow’s Economics subject area of the Adam Smith Business School. Successful scholarship applicants will be invited to apply for admission to the relevant PhD programme after they are selected for funding.