Covariance matrix prediction is a long-standing challenge in modern portfolio theory and quantitative finance. In this project, we investigate the effectiveness of Bayesian networks in predicting the covariance matrix of financial assets (specifically a subset of the S&P 500), evaluated against Heterogeneous Autoregressive (HAR) models. In particular, we consider both HAR-DRD, based on the DRD decomposition of the covariance matrix, and Graphical HAR (GHAR)-DRD, which is also based on DRD decomposition but also makes use of graphical relationships between the assets. To build the graph representing relationships between the assets, we apply Latent Dirichlet allocation (LDA) on the 10-K filings of each of the companies, and infer edges based on topic overlap. We show that this technique has limited usefulness in our setup, but provides recommendations on how it could be further improved based on our observations of its predictions.

Joint work with Kevin Minghan Li for the course project of 10-708 Probabilistic Graphical Models.


Link to our paper.