How XVA quants learned to stop worrying and trust the machine

Initial scepticism about using neural networks for derivatives pricing is giving way to enthusiasm

Jesper Andreasen knows a thing or two about speeding up the calculation of derivatives valuation adjustments, or XVAs.

So, when Microsoft claimed in a blog post last year that a neural network running on its Azure cloud platform could do the job up to 20 million times faster than existing models, the head of quant research at Saxo Bank – and self-styled Kwantfather – was sceptical.

Andreasen dismissed the idea on LinkedIn, saying it could only appeal to those with “more computers than brain cells”. Among other things, he argued the neural networks would take too long to train and may not be reliable enough to be used by the front office. Several other leading quants agreed.

Now, most of those questions have been put to rest. Last year, Scotiabank began using a deep neural network developed by Riskfuel, a fintech start-up, to approximate the outputs of the Monte Carlo models it uses for derivatives pricing. The neural network was trained within a day using the vast amounts of computing power available on Microsoft’s cloud platform. The bank credits the new system with improving both the speed and accuracy of its XVA calculations.

Since then, quants Antoine Savine and Brian Huge at Danske Bank have published research describing a novel way to speed up the training process. They trained their neural network using the sampled payoffs and sensitivities of derivatives, rather than examples of prices. This improves the quality of the training data while lowering the computational cost and time required to train the network. The new system has proven to be reliable and is already being used by the bank’s traders and risk managers. “Our research has resolved some key problems in the application of machine learning to risk,” Huge, Danske Bank’s chief quant analyst, tells Risk.net.

Others are now racing to catch up. At least four large dealers, including Barclays, Citi and HSBC, are actively exploring the use of neural networks for derivatives valuations. Ryan Ferguson, founder and chief executive of Riskfuel, reckons around a dozen banks are looking at implementing the technology.

Neural networks contain multiple hidden layers that transform the input data in ways that are often difficult to trace, let alone explain

The hurdles to wider adoption haven’t fallen away entirely, though. The training process can still be unduly burdensome for dealers with large trading books comprising a mix of complex products. Quants at JP Morgan have already decided to look for other ways to speed up XVA calculations.  

There’s also the lingering question of explainability. Neural networks contain multiple hidden layers that transform the input data in ways that are often difficult to trace, let alone explain. Increased regulatory scrutiny could slow the adoption of more complex deep learning techniques, such as those used for XVA calculations. A 2020 report from the European Banking Authority concluded that it was “premature to consider machine learning an appropriate tool for determining capital requirements, taking into account current limitations”, such as explainability.

Read the related feature here

Risk 0721 lead Eoin Coveney NB illustration
Eoin Coveney, www.nbillustration.co.uk

Danske’s neural network is not yet being used to report official numbers to regulators, although Huge has no doubt it will be.  

Other regulatory developments though may push banks to adopt neural networks. The standardised approach to credit valuation adjustments (SA-CVA), due to go live in January 2023, will require banks to calculate CVA sensitivities for all risk factors – a computationally intensive task that could push the limits of tried-and-tested technologies.

Among quants, there is a growing sense that deep neural networks will soon become a standard part of the XVA pricing toolkit. At Saxo Bank, Andreasen, who has been actively studying machine learning since 2017, is now considering using neural networks to model both the sensitivities of derivatives to underlying risk factors, as well as the co-movement of volatility surfaces. 

Editing by Helen Bartholomew

  • LinkedIn  
  • Save this article
  • Print this page  

Only users who have a paid subscription or are part of a corporate subscription are able to print or copy content.

To access these options, along with all other subscription benefits, please contact [email protected] or view our subscription options here: http://subscriptions.risk.net/subscribe

You are currently unable to copy this content. Please contact [email protected] to find out more.

You need to sign in to use this feature. If you don’t have a Risk.net account, please register for a trial.

Sign in
You are currently on corporate access.

To use this feature you will need an individual account. If you have one already please sign in.

Sign in.

Alternatively you can request an individual account here: