Skip to content
Menu
Technical University of Munich
Technical University of Munich

Workshop on Algorithmic Differentiation

 

Organizers

Viktor Mosenkis (RWTH Aachen, NAG) and Jacques du Toit (NAG)

 

Time and Venue

Monday, 9.9.2013: 10:00 – 13:00, Room 2.02.03

 

Description

The computation of sensitivities of pricing models with respect to input parameters (e.g. implied volatility, correlations) is a fundamental component of risk management.  Traditionally these sensitivities are approximated by finite difference quotients (so called “bumping”). This approach requires O(n) function evaluations and often results in infeasible run times for models with a large (or moderate) number of parameters.  Often such programs run for hours to produce the desired results. The adjoint model can reduce the computational effort for such problems by an order of magnitude. The method allows the computation of gradients the expense of O(1) function evaluations. In practical applications the factor lies typically between 2 and 5. Crucially, the computational effort is independent of n. The calculation of sensitivities, which previously had to be performed overnight, can now be done in a few minutes.

This workshop gives a thorough introduction to Algorithmic Differentiation (AD) which is used to build adjoint models.  First-order tangent-linear and adjoint methods are presented in detail.  The participants will learn how to use DCO (a software tool developed at RWTH Aachen) to differentiate their code.  We will also briefly show how DCO can be used to build higher order models.  A hands on session follows where participants apply DCO to a set of different codes, and closes with a discussion of a particular risk based application which demonstrates many issues in AD.

 

Registration

Please register for the workshop "Algorithmic Differentiation" using the form below. There is no additional registration fee for this workshop.

 

Agenda

  • Introduction to AD
    • First-order tangent-linear model
    • First-order adjoint model
    • Brief introduction to second- and higher order models
  • How to use DCO to produce
    • First-order tangent-linear models
    • First-order adjoint models
    • Brief introduction to obtaining higher order models
  • Hands-on session
    • Use DCO to compute sensitivities of a portfolio of different assets

  

To top