Exploring Quantum–AI for Financial Modelling: OQC, Citi and NQCC Collaborate on Derivative Pricing Innovation

The team first benchmarked a classical machine learning approach against analytical pricing models to establish a baseline. From there, quantum compression techniques were applied to parts of the neural network architecture. The results demonstrated that a significant reduction in model complexity could be achieved.

In one set of experiments, the approach substantially reduced the number of parameters in the neural network layer, while producing pricing results that remained broadly aligned with the original model in selected scenarios. While some increases in modelling error were observed depending on where compression was applied within the architecture, the results demonstrated a clear trade-off between model size and predictive performance: a key insight for designing future quantum-enhanced machine learning systems.

Just as importantly, the project highlighted current limitations, including that compression produced different results depending on which hidden layer of the neural network was compressed. Certain regions of the pricing problem proved more challenging for both classical and quantum approaches, particularly where pricing curves become mathematically sharp or highly sensitive to market conditions.

One of the most interesting outcomes of the project was the ability to significantly compress the size of a classical neural network using a quantum-based approach without dramatically affecting its predictive performance. By applying the Quantum-Train method to a shallow neural network architecture, we substantially reduced the number of parameters in a key layer. Despite this substantial reduction in model complexity, the pricing results remained broadly comparable to the classical model in many scenarios, with errors still within the same order of magnitude for several cases.

This highlights a promising direction for quantum-enhanced machine learning: using quantum representations to make large models more efficient while retaining much of their predictive power, an important step toward scalable AI systems for computationally demanding applications such as financial modelling. These findings provide valuable guidance for future research into hybrid quantum–classical modelling.