Spring Schedule, Imperial College London
 
The seminars in Statistics are held in 139 Huxley Building (http://www2.imperial.ac.uk/mathematics/about/contact.php ), at the South Kensington Campus at Imperial College London http://www3.imperial.ac.uk/contactsgettinghere/travelguidesandmaps/traveltoandfromimperialcollege , with coffee at 3 pm in the Mathematics common room. The seminars are open to a general audience, so please feel very welcome to attend.
 
The spring Statistics Seminars 2007 at Imperial College London are as follows:
 
19th January

 

14.00 Dr Tim Heaton, Department of Statistics, Oxford University,

A wavelet/lifting scheme based imputation method



It is often the case when performing a spatial survey that there exist sites of interest for which the underlying function is unobserved. In this talk we develop a new method to impute the value at these sites given information from neighbouring observed sites. Our approach takes advantage of the typically sparse representation of the underlying function in the wavelet domain within a Gibbs sampler framework.



We illustrate our method on both regularly spaced one-dimensional data with the classical wavelet transform as well as irregularly spaced two-dimensional data using a Voronoi based lifting scheme. In particular we consider the estimation of rainfall in the US where we compare performance of our wavelet technique to the now standard approaches of thin plate splines and kriging.

 

15.30 Ryan Anderson, Signal Processing Laboratory, Cambridge University

Using Phase Information from Decimated Complex Wavelets

 

This talk discusses methods and examples of using the phases of the Dual-Tree Complex Wavelet to perform edge detection and object matching across images. I will start by a review of complex-quadrature bandpass filters, such as Gabor filters, and mention general techniques for extracting infromation from their responses. We will then look at decimated complex wavelets, and see that the aforementioned techniques are inappropriate for decimated representations.

 

I will then introduce two new functions, the InterLevel Product and the SameLevel Product, which are able to extract useful structural information from these decimated coefficients. The most stable and informative coefficients may be clustered into sparse entities that we call Edge-Profile Clusters. We will then see how these clusters may be used in object matching scenarios.  

 

16th February

 

14.00 Dr Anastasia Papavasiliou, Department of Statistics, University of Warwick

Particle filters for non-ergodic systems

The goal is to construct a uniformly converging and efficient particle filter for non-ergodic system. First, I will discuss the asymptotic stability of the optimal filter, for such systems. Then, I will construct a particle filter that is guaranteed to converge uniformly. Finally, I will discuss ways of improving its efficiency. Note that these systems include the case where unknown parameters are treated as non-dynamic components of the state process.

 

15.30 Adam Johansen, Department of Mathematics, University of Bristol 

Rare Event Simulation with SMC Samplers

The problem of estimating ``rare event'' probabilities, and sampling trajectories of dynamic systems conditioned upon the occurence of rare events has received a great deal of attention in recent years. We present novel algorithms based upon sequential Monte Carlo techniques for addressing both of these problems for two broad classes of rare events.

Applications to a range of scenarios, and some results concerning the asymptotic behaviour of the systems which are proposed will also be presented.

 

23rd February

 

14.00 Darren Upton, Man Investments

Towards a time series model for the performance of hedge fund managers

 

To model the performance of hedge fund managers it is desirable to construct a time series model that will capture many features that are salient in financial datasets. These features include heavy tailed non-symmetric distributions, time varying variance, a mean that is a function of variance, and dependence between different managers. Each of these topics is briefly discussed with reference to a dataset which contains the performance of 4000 managers over the last 15 years. This leads us to propose a multivariate time series model for the dataset, which is fitted using maximum likelihood and accounts for  missing data.

 

9th March

 

14.00 Dr Michael Stumpf, Theoretical Genomics, Imperial College London

Sampling properties of random graphs and biological networks

 

15.30 Professor Glenn Shafer, Computer Learning Research Centre, Rutgers University

On Game Theoretic Probability

 

11th May

 

14.00 Professor Michael Wolf, Econometrics and Applied Statistics, Universität Zürich

Formalized Data Snooping Based on Generalized Error Rates

It is common in econometric applications that several hypothesis tests are carried out simultaneously. The problem then becomes how to decide which hypotheses to reject, accounting for the multitude of tests. The classical approach is to control the familywise error rate (FWE) which is the probability of one or more false rejections. But when the number of hypotheses under consideration is large, control of the FWE can become too demanding. As a result, the number of false hypotheses rejected may be small or even zero. This suggests replacing control of the FWE by a more liberal measure. To this end, we review a number of recent proposals from the statistical literature. We briefly discuss how these procedures apply to the general problem of model selection. A simulation study and two empirical applications illustrate the methods.

 

15.30 Professor Karim Abadir, Tanaka Business School, Imperial College London

Semiparametric estimation and inference for trending I(d) and related processes.

 

We deal with estimation and hypothesis testing in models allowing for trending processes that are possibly nonstationary, nonlinear, and non-Gaussian. Using semiparametric estimators, we obtain asymptotic confidence intervals for the trend and memory parameters, and we develop joint hypothesis testing for these. The confidence intervals are applicable for a wide class of processes, exhibit high coverage accuracy, and are easy to implement.

 
Kind regards,
Sofia