Graph-Based Factor Models for Interpretable Credit Spread Decomposition

Factor models are essential tools for understanding asset returns. Statistical factor models such as principal component analysis (PCA) and autoencoders have been widely used to reduce the high-dimensional panels of returns into a lower-dimensional latent space. Although effective at retaining much of the original variance, these models often lack inherent economic interpretation and rely solely on historical data, failing to incorporate contextual features such as asset characteristics into factor construction. Consequently, ad hoc analyses are often required to assign real-world meaning to latent factors. To address these limitations, this article introduces a novel graph factor model (GFM) that integrates domain-informed sparsity, explicitly connecting factors to financially validated features to enable interpretable and robust factor extraction. Extensive experiments on modeling corporate spread returns demonstrate that the GFM captures more variance, is more robust to missing data, and provides clearer economic insights than PCA, autoencoders, and instrumented PCA. By bridging the gap between statistical performance and economic interpretability, this new framework supports tasks such as performance attribution and offers valuable insights for portfolio management.