Exchange-traded financial products, such as stocks, treasuries and currencies, have benefited from a tremendous wave of technological innovation over the past 20 years, resulting in more efficient markets, lower reduced transactions and greater transparency for investors.
However, much of the capital markets has been left behind. The valuation of the instruments that make up the massive $500 trillion market for over-the-counter (OTC) derivatives – such as interest rate swaps, credit default swaps and structured products – has not the same degree of immediate clarity enjoyed by their simpler siblings.
In times of heightened volatility, traders and their managers need to know the impacts of market conditions on a given instrument throughout the day so they can take appropriate action. Reports reflecting conditions at the previous close of business are only useful in quiet markets and even then, companies with access to quick valuation and risk sensitivity calculations have a substantial market advantage.
Unlike exchange-traded instruments, where values can be observed each time the instrument is traded, the values of OTC derivatives must be calculated using complex financial models. The conventional way to achieve this is to use traditional Monte Carlo – a simple but computationally expensive probabilistic sweep through a range of scenarios and the resulting outcomes or finite difference analysis.
Banks spend tens of millions of dollars a year calculating the values of their OTC derivatives portfolios in overnight jackpots. These pesky parallel workloads have evolved straight from the days of the mainframe to run on on-premises clusters of conventional, CPU-bound workers, delivering a set of good results for any given day.
The use of conventional algorithms, real-time pricing and risk management are out of reach. But as the influence of machine learning extends to production workloads, a compelling pattern is emerging in scenarios and industries reliant on traditional simulation. Once computed, the output of traditional simulation can be used to train DNN models which can then be evaluated in near real time with the introduction of GPU acceleration.
We recently collaborated with Riskfuel, a startup developing rapid derivative models based on artificial intelligence (AI), to measure the performance obtained by running a model accelerated by Riskfuel on the Azure virtual machine instance ND40rs_v2 (NDv2 series) powered by NVIDIA. GPU versus traditional CPU-focused methods.
Riskfuel is a pioneer in using deep neural networks to learn the complex pricing functions used to price OTC derivatives. The financial instrument chosen for our study was the currency barrier option.
The first step in this trial was to generate a large pool of samples to use for training data. In this case, we used conventional CPU-based workers to generate 100,000,000 training samples by repeatedly running the traditional model with inputs covering the entire domain to be approximated by the Riskfuel model. The traditional model took an average of 2250 milliseconds (ms) to generate each rating. With the traditional model, the valuation time depends on the maturity of the trade.
The histogram in Figure 1 shows the distribution of evaluation times for a traditional model:
Figure 1: Distribution of valuation times for traditional models.
Once the Riskfuel model is trained, the valuation of individual transactions is much faster with an average of less than 3 ms, and no longer depends on the maturity of the transaction:
Figure 2: Riskfuel model demonstrating evaluation times with an average of less than 3 ms.
These results are for individual evaluations and do not utilize the massive parallelism that the Azure ND40rs_v2 VM can provide when saturated in a batch inference scenario. When called upon to evaluate trading portfolios, such as those found in a typical trading portfolio, the benefits are even greater. In our study, the combination of a Riskfuel-accelerated version of the Forex Barrier Option model and an Azure ND40rs_v2 virtual machine showed a performance improvement of more than 20 million times compared to the traditional model.
Figure 3 shows the throughput, measured in evaluations per second, of the traditional model running on an unaccelerated Azure VM compared to the Riskfuel model running on an Azure ND40rs_v2 VM (in blue):
Figure 3: Model comparison between the traditional running model and the Riskfuel model.
For wallets with 32,768 transactions, the throughput on an Azure ND40rs_v2 virtual machine is 915,000,000 ratings per second, while the traditional model running on CPU-based virtual machines has a throughput of only 32 ratings per second. This is a demonstrated improvement of over 28,000,000x.
It is essential to emphasize here that the acceleration resulting from the Riskfuel model does not sacrifice precision. In addition to being extremely fast, the Riskfuel model effectively matches the results generated by the traditional model, as shown in Figure 4:
Figure 4: Accuracy of the Riskfuel model.
These results clearly demonstrate the potential to supplant traditional on-premises high-performance computing (HPC) simulation workloads through a hybrid approach: using traditional cloud-based methods as the methodology to produce datasets used to train DNNs which can then evaluate the same set of works almost in real time.
The ND40rs_v2 Azure Virtual Machine is a new addition to the NVIDIA GPU-based Azure Virtual Machine family. These instances are designed to meet the needs of the most demanding GPU-accelerated AI, machine learning, simulation, and HPC workloads, and the decision to use the Azure ND40rs_v2 virtual machine was to take full advantage of the performance massive floating-point capabilities it offers to achieve the best batch-oriented performance for inference steps, as well as the highest possible throughput for model training.
The Azure ND40rs_v2 virtual machine is powered by eight NVIDIA V100 Tensor Core GPUs, each with 32 GB of GPU memory and with NVLink high-speed interconnects. When combined, these GPUs deliver a petaFLOPS of FP16 compute.
Riskfuel Founder and CEO Ryan Ferguson predicts that the combination of Riskfuel’s accelerated pricing models and NVIDIA GPU-powered VM instances on Azure will transform the OTC market:
“The current market volatility demonstrates the need for real-time pricing and risk management for OTC derivatives. The era of the night batch is coming to an end. And it’s not just the Azure ND40rs_v2 virtual machine’s lightning-fast inference that we love so much, but also the model training tasks. On this fast GPU instance, we reduced our training time from 48 hours to less than four hours! The reduced time to train the model, coupled with on-demand availability, maximizes the productivity of our AI engineering team. »
Scotiabank recently implemented Riskfuel models in its cutting-edge derivatives platform already live on the Azure GPU platform with NVIDIA GPU-powered Azure Virtual Machine instances. Karin Bergeron, Managing Director and Head of XVA Trading at Scotiabank, sees the benefits of Scotia’s new platform:
“By migrating to the cloud, we are able to create additional virtual machines if something requires additional scenario analysis. Previously, we did not have access to this type of on-demand computing. “Performance is welcome. This access to on-demand computing helps my team deliver better pricing to our customers.”