Neural networks and finance: derivatives pricing and risk management

What it is about

Use cases of neural networks in finance: generative adversarial networks (GANs) to create synthetic data applicable to market risk management and stress testing; variational autoencoders (VAEs) to analyze yield curve dynamics and their applications in stress testing; neural networks for pricing financial options; and new optimization algorithms for neural networks, such as gradient descent with fractional derivatives.

Session summary

The speaker, in the second of a series of webinars on artificial intelligence, starts from an explanation of what a neural network is: a complex function that captures non-linear relationships, with activation functions, several layers and training through backpropagation and gradient descent. He presents three applications. The first is generative adversarial networks to create synthetic data, useful when data are scarce, with an exercise on the Central Bank’s monthly economic activity index; he explains their advantages for stress testing compared with Monte Carlo simulations, and their limits: cost, complexity and the fact that training may not converge.

The second is variational autoencoders to reduce data dimensionality, which he contrasts with principal component analysis applied to the yield curve of Ministry of Finance securities, as a still-experimental non-linear complement. The third is his own research on optimization algorithms with fractional derivatives, which in an exercise valuing noisy European options reduced the number of iterations needed to train the network. In the questions he recommended using synthetic data as an additional tool, with care about quality, testing several architectures and checking that the generated data reflect the characteristics of the real data.

Topics covered

• What a neural network is and how it is trained
• Generative adversarial networks and synthetic data for stress testing
• Variational autoencoders versus principal component analysis
• Yield curve and dimensionality reduction
• Optimization with fractional derivatives and option valuation
• Limits: cost, complexity, convergence and data quality
• Questions: how to use these tools with limited data