Deep learning derivatives pricing

I made two simplistic TensorFlow (1.x) notebooks for the benefit of my students at Copenhagen University, to demonstrate how vanilla neural nets (deeply) learn pricing of European calls and high dimensional basket options, together with a comparison with conventional polynomial regression models (a la LSM) and a quick, simple introduction to the implementation of deep learning models in TensorFlow.

Notebook 1: European call in Black and Scholes

Notebook 2: Basket option in Bachelier

The notebooks run directly on Google colab, in the cloud and on GPU, no installation required.

The lecture slides and material are all found here: github.com/asavine/CompFinLecture

Some (much) more advanced considerations for efficiently learning prices of trading books, including “twin” neural nets who learn values and risks, and the super efficient differential regularization, are found here: slideshare.net/AntoineSavine/deep-analytics

Recorded workshop from Kings College London: AAD, Backpropagation and Machine Learning in Finance

Back in March, I gave a series of lectures at Kings College London on automatic adjoint differentiation, backpropagation and machine learning, and how it all connects and applies to risk management of financial derivatives.

The lectures were recorded and made freely available online, either from Kings own page:

https://nms.kcl.ac.uk/probability/workshopPages.php?id=8

or, on YouTube:

It is my understanding that the audience at Kings found the talks useful, and I hope that they may be beneficial to a wider audience on the internet.

QuantMinds 2019, Vienna

We had another fantastic conference from QuantMinds last week in Vienna, where I was honored to chair the Numerical and Computational Finance stream. The agenda was dominated by the application of machine learning to derivatives finance.

Brian Huge and I gave a talk on Deep Analytics to a vast audience of delegates, professionals and academics, including Bruno Dupire, Leif Andersen, Jesper Andreasen, Michael Dempster and many others. The room was full and some had to stand in the back for the whole 90 minutes!

Among many presentations focusing on big data, large infrastructures and deep neural networks, our talk stood out by a principled approach and the systematic integration and synergy of new methods from the deep learning world within the mature financial theoretical framework and in combination with other powerful financial techniques. Following Jesper Andreasen’s tradition of “computing XVA on iPad Mini”, we effectively trained small networks, within seconds to minutes, on trader workstations. We also proposed new techniques for the improvement of financial neural networks, not necessarily appropriate to other contexts, but particularly powerful for the pricing and revaluation of derivatives transactions and trading books. The most prominent of these techniques, “differential regularization”, leverages both AAD on the simulation side and back-prop on the neural side.

We extend billion thanks to our exceptional audience for a fantastic reception and feedback, to our colleagues of Danske Bank’s quantitative research for helping us polish the material and delivery, and to Danske Bank itself for encouraging and nurturing the research and development efforts necessary to properly produce game changing technologies like scripting, model hierarchies, parallel simulations, AAD and now, neural analytics.

It goes without saying that we heard other exceptional presentations, including, in my stream alone, Pierre Henry-Labordere‘s ground breaking work on the application of Schrodinger Bridges to stochastic volatility, and Jesper Andreasen’s rare and inspiring talk about scripting.

Outside of the stream, a series of filmed interviews gave me a chance to express my views on automatic differentiation (AAD), its exceptional contribution to risk management, its evolution and its future in quantitative finance combined with machine learning.

Finally, the presentation and interviews gave us an opportunity to introduce Danske Bank’s vision of “One Calculation Engine”. A core principle within Danske Bank’s quantitative research, this is a unique, single and consistent framework for the risk management of derivatives and the computation of regulations like XVA, CCR, FRTB… with the articulation of scripting, model hierarchies, parallel simulation engines, AAD and… deep analytics.

Antoine Savine, May 2019

Modern Computational Finance forum on Quora

A topic Modern Computational Finance was created on Quora for exchanging questions, answers, suggestions and comments about the Modern Computational Finance book, and, more generally, the topics of Numerical Finance, Monte-Carlo simulations, quasi-random number generators, parallel implementation in C++ and of course automatic differentiation.

https://www.quora.com/topic/Modern-Computational-Finance

out now

60th birthday: a tribute to Bruno Dupire

The RiO 2018 conference in mathematical finance, organized by IMPA, was held in Buzios, Rio de Janeiro, Brazil, on 24-28 November 2018, to celebrate the 60th birthday of Bruno Dupire, one of the most influential figures in the history of financial derivatives. The event gathered a number of leading scientists and professionals in one of the most productive and satisfying conferences of the year.

As one of his original alumni, I was given the responsibility to retrace some of Bruno’s ground breaking innovations, and, as a lecturer in Volatility, put them in perspective in the context of volatility modeling and derivatives trading.

See my slides here: