New numerical methods: How generative AI enhances quants’ toolkits
How is generative AI (GenAI) applied in quant finance? Youssef Elouerkhaoui, Managing Director, Global Head of Markets Quantitative Analysis at Citigroup, explores the challenges and implications.
In this interview, Youssef delves into the evolution of neural network topologies and the transformative power of the attention layer used in transformers. These advancements have expedited model training and enhanced pricing precision. Computational challenges still persist, but it’s clear that GenAI methods will be an important numerical methodology in quant finance moving forward.
The surprising impact of GenAI
In recent years, the exploration of GenAI methods, especially in deep pricing, has led to remarkable improvements. From the usage of neural networks years ago to the integration of Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs), the field has witnessed a dynamic evolution. However, the most significant leap has been the adoption of transformer models, which are instrumental for large language models like ChatGPT.
Transformers are essentially a neural network topology enhanced with an attention layer. This addition allows the network to prioritise important pieces of input data, mimicking human cognitive processes. By enabling a more efficient focus on essential data elements, transformers enhance the speed and precision of model training, resulting in more accurate pricing.
Ongoing challenges in GenAI methods
One significant hurdle is the implementation and performance of these models, particularly concerning computational resources. Training on GPUs is fast but complicated due to the need to handle large attention matrices. The transition between CPU and GPU can limit performance, emphasising the need for ongoing research into sparse attention and other optimisation techniques like KV caching.
Reshaping skills for next-generation quants
Youssef argues that GenAI should be considered another tool in the broader set of numerical methods, including traditional techniques like Monte-Carlo simulations and Partial Differential Equations (PDEs). The expectation moving forward is for quants to integrate these AI-driven methods into their day-to-day analyses seamlessly, enhancing their roles without considering them as separate applications.
Discover the future of quant finance with leading practitioners and experts at QuantMinds International. Save the date – 16-19 November 2026!
