The Statistics Section has an international reputation for conducting methodological and applied statistical research at the highest level. Particular areas of current activity include statistical genetics and biostatistics, statistical methods in retail financial services, time series, core statistical methodology, classification and data mining, with many interactions and overlaps between these areas of research.
Statistical Genetics, Bioinformatics and Biostatistics
Research in biostatistics, statistical genetics and bioinformatics ranges from structured modelling of high dimensional 'omics data to systems biology. The aim is to apply statistical methodology and structural modelling to data generated by next-generation high-throughput technologies including ancillary information available from molecular biology in order to identify potential targets for therapeutic intervention.
The goal is achieved by applying modern statistical methods (unsupervised pattern discovery, sparsity models, hierarchical Bayesian models) and computational intensive inference (permutation procedures, MCMC, SMC) to exploit causes of human genetic variation in final phenotypes and integrate them with other intermediate forms of phenotypic variation such as gene expression, metabolites and
Statistics in Retail Financial Services
The aims and objectives of the Statistical Methods in Retail Financial Services Research Group are:
To apply statistical methods in the financial sector of industry; and
To develop new statistical methodology arising from the novel challenges presented by these applications.
It is very apparent that the level of sophistication of the methods and tools used by the banks to control their retail credit operations is increasing rapidly, and we aim to continue to be in the vanguard of this development. Our research involves all aspects of mathematics and statistics relating to the retail finance industry, including fraud detection, portfolio modelling, default correlation and risk management.
We have numerous links with organisations within the retail banking sector, and are always involved in discussions with banks and other financial bodies about possible collaborations or research sponsorships. Previous projects have been sponsored by Fair Isaac, Link Financial, and other bodies.
Time Series, Spatial Statistics and Signal Processing
This theme covers research in the broad area of time series from novel stochastic processes to matrix-valued wavelets, with application areas ranging from finance and energy planning to the analysis of geophysical flows influenced by the rotation of the Earth. The group also pursues research on the areas of continuous time modelling of time series, statistical methods for stochastic processes and financial econometrics (In particular, stochastic volatility modelling and estimation; High frequency financial data)
Statistical Theory and Applied Probability
The Statistical Theory and Applied Probability group is active in the development of new statistical methodologies, as well as in adapting, extending and determining the theoretical properties of existing methods and computational algorithms. Research interests within the group include new advances in the theory and methodology of Monte Carlo methods. Moreover, research in applied probability focuses also on (spatio-) temporal stochastic modelling and inference as well as on applications in financial mathematics.
The Statistics Section at Imperial College London has close ties with the Astrophysics Group, and members of both groups make up the recently-formed Imperial Centre for Inference and Cosmology (ICIC). We are involved in researching problems of astrostatistics, an exciting and emerging field devoted to the application of modern statistical techniques to solve problems of astronomical data analysis and to the use of astronomical data-sets to test and develop new statistical methods. The astrostatistics research at Imperial has two main threads. The main thread of astrostatisics research here is the development - and application - of principled statistical methods to make sense of the massive data-sets that are now routinely produced by large astronomical surveys. This can include fundamental techniques for data processing (or "data reduction" as it is often termed in astronomy), as well as science exploitation projects, such as rare object searches (e.g., looking for the most distant objects in the Universe) and Bayesian parameter inference (e.g., measuring the ages and other properties of stars given only very imprecise measurements). We also use astronomical data as a testing ground for innovative statistical techniques, with a particular emphasis on examining the roles of priors in Bayesian parameter estimation and model comparison. The latter is particular important in cosmology, as while the available data are exquisitely characterised and the theoretical models are entirely quantitative, the lack of well-motivated priors on model parameters makes model-testing and checking problematic. We are actively involved in finding new methods to solve such problems.
Bayesian Methods and Computation
Bayesian methods offers a consistent way of characterising un certainty in statistical inference while incorporating multiple sources of information, perhaps including prior information. In many cases such methods are computationally demanding. The work carried out within the Bayesian Methods and Computation theme focuses on t he development and application of highly efficient approaches to performing Bayesian inference. This involves the development of state-of-the-art methodology, as well as tackling the computational issues that arise from their practical implementation, in particular by using efficient computational methods and/or large-scale cluster computing and other highly parallelised hardware.
Much of the research within this theme is directly motivated by real-world applications, in finance, retail and engineering, and by important scientific questions arising in astronomy and the medical and life sciences. Our work is therefore driven through strong links with our many scientific and industrial collaborators, and our members are frequently invited to give high profile seminars and talks at leading institutions, academic societies and companies throughout the world.
Big Data and Statistical Machine Learning
It is now a recognised fact that we are facing a data revolution in both sciences and industry. In part, this represents a renewed interest in existing areas such as Data Mining and Business Intelligence, which have traditionally been handled using classification and machine learning technology. However, recent emphasis on Big Data also encompasses new problems, driven by infrastructural advances giving rise to databases of unprecedented scale (e.g., distributed databases or data streams), as well as altogether new data formats (e.g., free-form text, networks, etc.). The Machine Learning and Big Data group focuses on methodological research, as well as the interface between statistical methodology and cutting-edge infrastructure (such as Hadoop and stream processing languages). The group is active in modernising practice in existing areas such as fraud detection and retail credit; and on emerging applications, as well as emerging applications such as cybersecurity and neuroimaging.
(Adams, Anagnostopoulos, Hand, Heard, Montana)
There are many security contexts in which there is a requirement to monitor the evolution of a large, dynamic network. Such networks might be telecommunications between individuals within communities of interest, or data traffic flows on computer networks holding sensitive information. This research group is interested in developing statistical methodology for analysing the evolution of such networks, with a strong focus on computationally scalable methods. A principle aim is enabling on-line anomaly detection, so that any intruders or abnormalities in the network can be quickly identified in real time.
The group has active collaborations within Imperial and with external partners both in the UK and the US.