An inhomogeneous Poisson process model for count time-series data

Event Date: 

Wednesday, November 13, 2024 - 3:30pm to 4:30pm

Event Date Details: 

Wednesday November 13th, 2024

Event Location: 

  • PSYCH 1902

Event Price: 

FREE

Event Contact: 

Dr. Ioannis Chalkiadakis

  • Department Seminar
Abstract:
This talk will present a novel time-series framework for leveraging count time-series data that are observed at regular or irregular time intervals.
We will begin by summarising a recently developed framework for leveraging text time-series data in econometrics and political sciences. We will start with what we believe constitutes challenges in Natural Language Processing for statistics, econometrics and political sciences research, before presenting the statistical framework for text time-series and sentiment signals construction. We will then present our model for count time-series data, which consists of an inhomogeneous Poisson process with a chi squared process intensity function, driven by text-based covariates. We demonstrate an efficient optimisation procedure leveraging control points with a variational inference framework. An example of the model's utility in capturing statistical signatures of texts will be illustrated via some preliminary results in the context of political science and American presidential rhetoric.