Learning Relational Asset Embeddings from Financial Returns Time Series

Financial markets are complex systems that play a crucial role in the global economy. Modelling and analysing these markets is challenging due in part to the intricate relationships and interactions between financial assets. With most existing research applying machine learning to financial markets focusing on returns forecasting for individual assets, this thesis addresses a gap in the literature by developing novel frameworks that learn relational embedding representations of financial assets from their returns time series data.Inspired by recent advancements in representation learning, we propose methods that leverage pairwise similarity information and self-supervised learning objectives to uncover the latent structure of financial markets. Our contributions include new time series similarity metrics, machine learning algorithms for representation learning, and methods for mining relational samples from time series data. We introduce a hybrid similarity metric tailored to financial time series, extend matrix factorization techniques to learn asset representations from similarity matrices, and propose a classification-based embedding learning framework. We also develop a contrastive learning framework to generate informative embedding representations of financial assets, proposing novel loss functions and sampling strategies.The effectiveness of our methods is primarily demonstrated through applications in two financial tasks: portfolio optimization and market sector classification. Our approaches lead to statistically significant performance improvements compared to benchmark models, highlighting the practical value of our research in addressing real-world financial problems. The proposed methods have the potential to augment the way financial institutions currently model and analyse markets, ultimately leading to more informed decision-making and improved risk management.