Financed Informed Auto-Encoder Market Models

Specialeforsvar: Mathilde Freja Scavenius

Titel: Financed Informed Auto-Encoder Market Models

Abstract: In this thesis we attempt to describe the dynamics of yield curves by applying neural networks, in particular, auto-encoder models. To this end, we develop four main models - one to show that this method is indeed viable, another to investigate the generative potential of auto-encoder models and finally both of these models are extended with a condition of
no arbitrage. The generative potential is investigated by developing so-called variational auto-encoders, which are essentially probabilistic versions of regular auto-encoders. The condition of no arbitrage is put forth by modifying the loss function of the models, for which we use financial arbitrage theory. The implementation of these models leads to numerous interesting findings. While the models perform well using only two latent factors, a lot is gained from adding a third factor, specifically the ability to capture inverted curves. We also find that variational auto-encoders do in fact seem to better generate new data as compared to the regular auto-encoders. These results underscore the immense potential of neural network-based models in accurately modeling yield curve dynamics.

Vejleder: Rolf Poulsen
Censor:    David Sloth Pedersen, Danske Bank