Quant house of the year: Credit Suisse

Asia Risk Awards 2019

Photo of George Hong
George Hong, Credit Suisse

Callable equity-forex hybrid range accrual notes are particularly popular with investors in Asia who seek alternatives to plain-vanilla low-return investment products.

The equity-forex hybrids are structured notes with coupons paying an amount linked to performance of an equity underlying, usually an index, and a foreign exchange rate typically versus US dollars. The investor receives the coupon depending on whether the two underlyings’ values lie within a predefined range, although the dealer has the option to terminate the note before expiry.

They are attractive because the coupon rate is much higher than that of standard rate products, potentially reaching 5% or 6% in a low rates regime like the current one.

Dealers in the area are well aware of the opportunity of this market and race to meet the demand, which comes primarily from private banks and institutional investors.

These notes are, however, notoriously complex to value and risk manage, because of the high dimensionality of the problem, their embedded optionality and the path dependency of the payoff, among other reasons. On top of that, banks need to perform an increasing number of simulations to satisfy both internal needs for scenario analysis and requirements of risk calculations from regulators.

The quant team of Credit Suisse came up with a multi-faceted solution that includes four separate but complementary innovations on the management of hybrid range accruals.

“The trading desks and the senior management had constantly emphasised to the quant team how critical an optimised analytics platform is to the overall profitability and successful risk management of this business line,” reveals George Hong, managing director and the head of Asia-Pacific quantitative and risk strategies in Hong Kong.

The standard approach for managing these products is to implement a method based on partial differential equations, but that struggles to deal with the number of underlyings and stochastic factors, and to include features such as redemption or notional protection.

Credit Suisse’s approach is instead to optimise an American Monte Carlo (MC) algorithm, robust and efficient enough to handle trading and control functions. This is supported by the four solutions that have been deployed progressively as they were individually approved by the internal validation team.

Firstly, the key innovation on the modelling side is on the treatment of the quanto skew, which is crucial to the pricing of the equity-forex hybrid range accrual. For long, it has been market practice to adjust the volatility of the skew term in the pricing formula with a constant.

However, this is no longer sufficient. Hong himself developed an alternative solution that uses stochastic collocation techniques, which improves the accuracy of the pricing while keeping the framework relatively simple. The model is detailed in a paper Hong published in Risk.net earlier this year and discussed in a podcast

The trading desks and the senior management had constantly emphasised to the quant team how critical an optimised analytics platform is to the overall profitability and successful risk management of this business line
George Hong, Credit Suisses

The MC algorithm itself is at the core of the framework and, especially for exotic products like the hybrid range accrual, it requires enormous computational resources. Banks need to invest heavily in CPU units or add more powerful processors like graphics processing units (GPU) to the farm. However, the former is costly and the latter is not as straightforward as it may sound. Using GPUs means rewriting codes and adapting data libraries, which take time and money for training and maintenance. GPUs are not suitable to perform all tasks in the procedure, so they will always co-exist with CPUs.

This led to a global effort within Credit Suisse in order to find a way to integrate the two technologies efficiently. The concept-of-new-data type, which was called field, was introduced as a tool that is compatible with both GPU and CPU.

“Quants don’t need to learn about how GPU works or its programming language, such as CUDA, because the interface enables the code to look the same to them in both the cases of programming for GPU and for CPU,” explains Hong.

Jean-Jacques Fabre, head of risk platform, Apac, and David Wilkinson, senior director in the Apac quant team, drove the GPU project from the design stage and code implementation phase, all the way to taking the trading books live globally, working with IT and quant teams across different regions. This solution has been developed not just for hybrid range accruals, but for much wider application and it is designed to account for specific requirements from different trading desks from the onset. The result was an average computing speed up of a factor of a hundred.

Then came the optionality. Isolating and pricing the embedded call option in a range accrual product is no easy task. The bank has the right to terminate (call) the note early at predefined times, just like a Bermudan option. Typically, Bermudan options are priced using a least square regression algorithm, using polynomials as basis functions.

While this works well for most products, polynomials are not best-placed to describe the payoff of range accruals. Hong’s team therefore adapted the radial basis function neural network (RBFNN), a machine learning tool, to approximate the exercise boundary of the option. RBFNN is fast to train and has proven to be more robust than the polynomial approach.

Another improvable process is related to the coupon. The MC algorithm needs to simulate paths while verifying whether the underlyings are within their range, and price taking that into account. This can be computationally expensive and slow the process down significantly.

The solution Credit Suisse put in place is instead that of using an analytical approximation to interpolate the range probability term structure. In other words, not all time points are considered for possible violation of the range, but only selected few, which are representative of the probability term structure. This turns out to be much faster than the brute force daily sampling approach and, perhaps counterintuitively, more accurate than the standard approach with regards to the calculation of prices and risk sensitivities.

  • 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: