Fri, 27 May 2016

13:00 - 14:30
L6

Deep Learning for Modeling Financial Data

Justin Sirignano, postdoc at Imperial College.
(Imperial College London)
Abstract
Deep learning has emerged as one of the forefront areas in machine learning, achieving major success in imaging, speech recognition, and natural language processing. We apply deep learning to two areas in finance: (1) mortgage delinquency and prepayment and (2) limit order books. Using datasets unprecedented in size, we show that deep neural networks outperform several status quo approaches. Due to the heavy computational cost from both the size of the models and the data, we use GPU clusters to train the models.
Fri, 20 May 2016

13:00 - 14:30
L6

Talks by Phd Students

Our Phd Students Wei Fang and Alexander Vervuurt
(Mathematical Insitute, Oxford)
Abstract

Wei Title: Adaptive timestep Methods for non-globally Lipschitz SDEs

Wei Abstract: Explicit Euler and Milstein methods are two common ways to simulate the numerical solutions of
SDEs for its computability and implementability, but they require global Lipschitz continuity on both
drift and diffusion coefficients. By assuming the boundedness of the p-th moments of exact solution
and numerical solution, strong convergence of the Euler-type schemes for locally Lipschitz drift has been
proved in [HMS02], including the implicit Euler method and the semi-implicit Euler method. However,
except for some special cases, implicit-type Euler method requires additional computational cost, which
is very inefficient in practice. Explicit Euler method then is shown to be divergent in [HJK11] for non-
Lipschitz drift. Explicit tamed Euler method proposed in [HJK + 12], shows the strong convergence for the
one-sided Lipschitz condition with at most polynomial growth and it is also extended to tamed Milstein
method in [WG13]. In this paper, we propose a new adaptive timestep Euler method, which shows the
strong convergence under locally Lipschitz drift and gains the standard convergence order under one-sided
Lipschitz condition with at most polynomial growth. Numerical experiments also demonstrate a better
performance of our scheme, especially for large initial value and high dimensions, by comparing the mean
square error with respect to the runtime. In addition, we extend this adaptive scheme to Milstein method
and get a higher order strong convergence with commutative noise.

 

Alexander Title: Functionally-generated portfolios and optimal transport

Alexander Abstract: I will showcase some ongoing research, in which I try to make links between the class of functionally-generated portfolios from Stochastic Portfolio Theory, and certain optimal transport problems.

Fri, 06 May 2016

13:00 - 14:30
L6

Some remarks on functionally generated portfolios

Johannes Ruf
(UCL)
Abstract

In the first part of the talk I will review Bob Fernholz' theory of functionally generated portfolios. In the second part I will discuss questions related to the existence of short-term arbitrage opportunities.
This is joint work with Bob Fernholz and Ioannis Karatzas

Dynamic Bayesian networks as formal abstractions of structured stochastic processes
Soudjani, S Abate, A Majumdar, R Leibniz International Proceedings in Informatics Lipics volume 42 169-183 (01 Aug 2015)
The number of accessible paths in the hypercube
Berestycki, J Brunet, E Shi, Z Bernoulli volume 22 issue 2 653-680 (01 May 2016)
Thu, 03 Mar 2016
11:00
C5

'Additive extensions and Pell's equation in polynomials'.

H.Schmidt
(Oxford)
Abstract

We will discuss families of Pell's equation in polynomials 
with one complex parameter. In particular the relation between 
the generic equation and its specializations. Our emphasis will
be on families with a triple zero. Then additive extensions enter 
the picture. 

Thu, 24 Nov 2016

16:00 - 17:00
L3

An engineer's dive into Oxford Applied Maths, and becoming faculty at a Medical School

Athanasios Tsanas
(University of Oxford)
Abstract

In this talk, I am reflecting on the last 8 extremely enjoyable years I spent in the department (DPhil, OCIAM, 2008-2012, post-doc, WCMB, 2012-2016). My story is a little unusual: coming from an Engineering undergraduate background, spending 8 years in the Maths department, and now moving to a faculty position at the Medical School. However, I think it highlights well the enormous breadth and applicability of mathematics beyond traditional disciplinary boundaries. I will discuss different projects during my time in Oxford, focusing on time-series, signal processing, and statistical machine learning methods, with diverse applications in real-world problems.

Thu, 17 Nov 2016

16:00 - 17:00
L3

Modelling Anti-Surfactants and Thixotropic Lubrication

Stephen Wilson
(University of Strathclyde)
Abstract

In the first part of the talk, I will describe a fluid-dynamical model for a "anti-surfactant" solution (such as salt dissolved in water) whose surface tension is an increasing function of bulk solvent concentration. In particular, I will show that this model is consistent with the standard model for surfactants, and predicts a novel instability for anti-surfactants not present for surfactants. Some further details are given in the recent paper by Conn et al. Phys. Rev. E 93 043121 (2016).

 

In the second part of the talk, I will formulate and analyse the governing equations for the flow of a thixotropic or antithixotropic fluid in a slowly varying channel. These equations are equivalent to the equations of classical lubrication theory for a Newtonian fluid, but incorporate the evolving microstructure of the fluid, described in terms of a scalar structure parameter. If time permits, I will seek draw some conclusions relevant to thixotropic flow in porous media. Some further details are given in the forthcoming paper by Pritchard et al. to appear in J Non-Newt. Fluid Mech (2016).

Subscribe to