PhD students

Forecasting

PhD students Zane and Cixiao are currently working with me and Prof. Niall MacKay on statistical methodology for making forecasts given information and opinion from experts. This differs slightly from more established methodology developed for making forecasts given physical measurements from sensors or machines. A key difference is that the distributions of ‘measurement errors’ from machines tend to be fairly simple whereas the errors made by groups of people can be highly complex.

Zane Hassoun

Zane is currently looking at crowd-sourcing numeric forecasts and aggregating them in clever ways. More specifically, he is developing methodology for quickly detecting changes in forecasts as time progresses and new information arrives.

Cixiao Jiang

Cixiao is looking at textual arguments that accompany numeric forecasts. She is investigating how the text might be used to inform the aggregation of the numbers.

Sports

India and Tara are working with me and Dr. Jess Hargreaves on statistical methodology with application to sports data. Sports are obviously interesting in their own right, but they also produce a lot of data, full of strange signals and artefacts, that we can use as a test-bed for models that are eventually used for other things.

India Richmond

India is working on models for capturing complex interactions in high-dimensional distributions. She tests her models mainly on data sets from sports matches, which are characterized by a variety of features that prove challenging for conventional statistical methods.

Tara Broughton

Tara’s work involves measuring the differences between data sets and models, both being understood as (possibly atomic) distributions. These differences have interesting implications if, for example, we consider augmenting a small data set with a larger one that is related but not actually from the distribution for which we want to make predictions or inferences.

Neural nets

Junxuan is supervised by me and Prof. Marina Knight. With this work we are making connections between statistical ideas and techniques developed in the 20th century, typically for use with small data sets resulting from designed experiments, and those developed in the last decade, typically designed to handle huge amounts of data from the internet or autonomous sensing devices.

Junxuan Jiang

Junxuan is working on the regularization of neural networks. This topic is particularly interesting for the way it relates the generalization abilities of models with the limits placed on the computational resource they can access.