Murex machine learning research captures specifics of complex derivatives pricing

Regulatory demands have pushed quant teams to evolve their models, and in order to stay competitive, the application of machine learning techniques are becoming ever more important. Do models perform well enough? Murex shares their recent research results in derivatives pricing, using machine learning techniques.

Training neural networks to replicate a derivative pricing model with supervised learning represents a tremendous opportunity in financial services.

Conscious of this opportunity, Murex has worked to create machine learning architecture that captures the specifics of sophisticated financial product and model dynamics.

Well-designed and well-trained, Murex’s neural network produces very accurate prices along with their sensitivities.

An in-depth white paper outlines the Murex approach.

Murex experts examined the optimal make up of neural network architecture. They also discovered how it must be trained to attain a high degree of precision and reliability in trading and risk management applications.

A bit of background

In the context of pricing derivatives products, computational demand has surged in previous decades. The consensus is that the demand will expand continuously.

The growing complexity of quantitative models; the regulatory demands for broad stress testing and scenario analysis (e.g., under Fundamental Review of the Trading Book, or FRTB); and the emergence of central desks dedicated to the active hedging of valuation adjustments (XVA) all factor into the requirements of a diverse body of financial institutions.

In response to these challenges, the industry has adapted and continuously innovated, employing new technologies and computational strategies. These innovations often fall short, Murex experts write.

Major investments in large-scale infrastructure required to support computational needs can be prohibitive. Performance gains may not justify the costs. Homing in on XVA, the computational demand of Monte Carlo simulations—often requiring hundreds of millions of evaluations—makes the use of the most sophisticated trading models impractical, if not impossible, for intraday or daily batch calculations. Consequently, simpler models are often implemented.

Training neural networks to replicate a derivative pricing model using supervised learning presents an opportunity to address these challenges. The paper describes an ambitious project to replicate an existing, mathematically tractable model with very high fidelity by training Murex’s machine learning model on high dimensional data, including financial derivatives contracts data and term structure data, such as yield curves or volatility surfaces.

It is tempting to consider a “generic” neural network taking product characteristics, calibrated model parameters and market data as inputs, and then, producing an approximated price as an output. This type of implementation only partially succeeds in providing precise and reliable results.

Murex experts aimed to design a fit-for-purpose machine learning architecture highly tailored to the specifics of the derivative product and pricing model dynamics.

To accomplish this, Murex’s approach decomposes the evaluation process into a series of sequences. Each sequence is meticulously designed to capture significant contractual events and receive the most pertinent inputs. This methodological segmentation ensures that the model’s attention is laser-focused on relevant data.

A fit-for-purpose neural network makes it possible to reach a high degree of precision when approximating prices of derivatives.

The use of such models opens new avenues for utilizing the most advanced models in contexts demanding massive numbers of evaluations under large sets of scenarios, such as in XVA or in potential future exposure, in batch or in real time. It also opens the door to building even more sophisticated models for trading or risk management—models that would previously be deemed impractical due to their computational intensity.

Additionally, this approach leads to a significant decrease in the need for extensive computational resources, which reduces both financial cost and environmental impact.

Meet the Murex team at QuantMinds International this November.