The Young Researcher's meeting is a fortnightly meeting for postgraduate students and postdocs. It provides an informal forum where we discuss research, exchange ideas and learn from and with each other. We usually start the meeting with a seminar on topics from or related to statistics, which is then followed by a discussion over coffee and biscuits.

Recent topics include; generalised linear mixed models, survival analysis, meta-analysis, probabilistic expert systems, image analysis and statistics for insurance.

Everyone is welcome to attend. All meetings take place during term time in the Statistics Common Room at 1600 unless otherwise advertised. The current organizer of this seminar is Flavio Goncalves.

Spring, 2009/10

Krzysztof Latuszynski - 02/03/2010

Making black boxes out of black boxes - one year later

In his opening talk for 2008/9 academic year YRMs professor Gareth Roberts advertised the problem of designing a black box that outputs a 2p-coin  (i.e. one that gives heads with probability 2p and tails with prob 1-2p) using a given black box that outputs a  p-coin. We have now solved the probelm!
This is joint work with  Ioannis Kosmidis, Omiros Papaspiliopoulos and Gareth Roberts.

Duy Pham - 16/02/2010

Measuring vega risks of Bermudan swaptions under the Markov-Functional model

Markov-Functional (MF) models form a popular class of models in which the value of pure discount bonds can be expressed as a functional of some (low-dimensional) Markov process. We shall consider a particular application of MF model, the Bermudan swaptions which are by far the most common in the interest rate derivatives market. Practically, calculation of risk sensitivities for a Bermudan swaption is as important as calculation of its value. In this work, we consider different parametrizations of the driving Markov process and their implications on the Bermudan swaption's vega risks.

Ashley Ford - 02/02/2010

Indian Buffet Epidemics

After an intoduction to existing models for Epidemics and the motivation for a new model.
The Indian Buffet Process is described and their combination to give Indian Buffet Epidemics is
described.

Chris Jewell - 19/01/2010

Busting through your calculations: an introduction to Buster, the departmental high-performance computing cluster

Modern statistics is increasingly requiring the use of high
performance computing as datasets get larger, and algorithms get more
complex.  Buster, the departmental computing cluster, is designed to
help you power through your calculations by providing fast processors
with access to large amounts of both RAM and disc space.  It is intended
to be a safe environment for research development, and uses many of the
features found on the larger supercomputers such as those found in
CSC.   This talk will give an introduction to using Buster and what it
can do to enhance your research.  Topics will include how to log in,
where to store your files, how to run your algorithms, and brief
overview of the software currently available.  All are welcome to attend
this introduction, and there will be a "signing in" session at the end
for those of you who wish to have an account on Buster, but don't
currently have one.

Autumn, 2009/10

Maria Costa - 08/12/2009

Single and Double Penalty Spline Regression Models with Applications

Penalized spline regression models are a popular statistical tool for curve fitting problems due to their flexibility and computational efficiency. In particular, penalized cubic spline functions have received a great deal of attention. Cubic splines have good numerical properties and have proven extremely useful in a variety of applications. Typically, splines are represented as linear combinations of basis functions. However, such representations can lack numerical stability or be difficult to manipulate analytically.
We propose a different parametrization for cubic spline functions that is intuitive and simple to implement. Moreover, integral based penalty functionals have simple interpretable expressions in terms of the components of the parametrization. Also, the curvature of the function is not constrained to be continuous everywhere on its domain, which adds flexibility to the fitting process. We consider not only models where smoothness is imposed by means of a single penalty functional, but also a generalization where a combination of different measures of roughness is built in order to specify the adequate limit of shrinkage for the problem at hand. The proposed methodology is illustrated in two distinct regression settings.

Peter Windridge - 24/11/2009

Recursive ternary majority revisited

I shall discuss results and open questions related to the
following problem:  take a complete ternary tree of depth n (i.e. 3^n
leaf nodes and all other nodes have three children) and assign
independent bernoulli(1/2) random variables to each of the leaves.  The
value of the other nodes is defined recursively to be the majority value
of its children (of which there are three, so there is always a
majority).  The only way we can find out about the tree's values is by
paying to look at the values of the leaves, with each observation
costing £1.

What is the cheapest way to find out the value of the root?

Thais Fonseca - 10/11/2009

Measuring separability of spatiotemporal covariance functions

In this work, we construct a measure of space-time dependence for general
nonseparable (possibly nonstationary)  covariance models. It is well known
that nonseparable covariance functions are more realistic for modeling
many geophysical and environmental processes. However, little is known
about the strength of dependence in space-time that is achieved by the
models proposed in the recent literature. We compute the proposed measure
for various nonseparable models and we show that some of them generate a
very limited range of nonseparability in space-time. Moreover, we
illustrate that certain space-time interaction parameters might have a
non-monotonous relation to our measure of separability, and they might not
be the only parameters affecting the degree of nonseparability obtained by
the model.

Guy Freeman - 27/10/2009

Using dynamic staged trees for discrete time series data: robust prediction, model selection and causal analysis

A new graphical model is proposed for discrete-valued discrete-time
data. We define the dynamic staged tree and implement a
one-step ahead prediction algorithm using multi-process modelling
and the power steady model that is robust yet also easy to
interpret. We also demonstrate how to analyse causal hypotheses on
this model class. We illustrate our techniques with a real educational
example.

Come and join us!