Main Speakers and courses

Title: Combining Features, Kernels and Algorithms
Abstract: Recently various methods have been proposed to combine multiple learners to improve accuracy. Combining learners is useful only when the learners are complementary and to achieve diversity, it has been
proposed to combine (1) features (from different representations/modalities/sources), (2) kernels (using different
measures of similarity), and/or (3) learning algorithms (with different inductive bias). We will discuss these different methods together with experimental results.

Title:   Convex measures of risk
Abstract:  Monetary measures of risk  quantify  minimal capital reserves, that should be maintained by financial institutions  in order to balance undertaken risks  and ensure financial stability. In this course I will focus on convex monetary risk measures. First, I will give an overview of the theory in the static case, explaining the axiomatic approach and providing the robust representation of convex risk measures. Then I will discuss risk measures in the dynamic environment, where risk assessment is updated over the time in accordance with the new information. In particular, I will characterize various time consistency aspects of dynamic risk assessment, and I will show how uncertainty about the time value of money can be taken into account. The results will be illustrated by examples. The course will be based on the book "Stochastic Finance" by H. Foellmer and A. Schied, and on joint works with Hans Foellmer and Beatrice Acciaio. 

Title: Online Learning in Data Analysis
Abstract: The old days of processing data off line are over. Most of the recent interesting applications of controls and machine learning in the manufacturing, service and entertainment industries require instant processing of streaming data, i.e. processing one sample at a time. It turns out that the field of adaptive signal processing has develop a large class of algorithms solving optimally the approximation problem for the linear and the nonlinear model using stochastic gradient approximations. This will be the focus of the course.
We will cover the following topics
1-      Filtering versus regression, and the importance of time
2-      Solving least squares with search algorithms: adaptive filtering
3-      Affine projection Algorithms
4-      Neural Networks for nonlinear regression and classification
5-      Reproducing kernel Hilbert space algorithms for regression and classification

Title: Harnessing the Fat Tails - The Case of Operational Risk Management and Measurement
Abstract: One stylized fact of financial markets contents that the probability distribution functions describing market data are fat-tailed. Nevertheless, most of the risk-management instruments rely directly or indirectly on normal distributions. The loss distributions of operational incidents in banks are both heavily skewed and extremely fat-tailed. Operational risk is defined as the risk of loss due to inappropriate or failed infrastructure, processes, human resources, or due to external impact.
In my talks, I will discuss the challenges of measuring operational risk in banks, and how they were mastered in a network of 430 independent savings banks in Germany. In addition to the collection of loss information in each bank, this involves the use of scenario analysis and the operation of central data pools. I will describe in detail the development and calibration of a risk measurement tool which accurately models and aggregates the fat-tailed distributions found from almost 50,000 loss data from the pool.
Tentative outline:
Part 1:
Part 2:
Part 3: