Towards Representation Learning for Cross-Sectional Portfolio Construction

Traditional cross-sectional trading heuristics summarize an asset’s return history using fixed, hand-crafted weighting schemes, implicitly imposing strong assumptions on how past returns should be aggregated. Such summaries collapse rich structure in return paths into a single scalar signal, potentially discarding structure that is useful for downstream portfolio objectives such as risk-adjusted performance. We propose a representation learning approach that learns task-specific embeddings of return paths directly from raw past returns. Using a lightweight temporal convolutional network, each asset’s recent return history is mapped to a low-dimensional representation, which is aggregated and converted into a dollar-neutral long--short portfolio. The model is trained end-to-end to optimize an out-of-sample risk-adjusted portfolio objective and empirically yields promising risk-adjusted returns, outperforming traditional momentum and reversal heuristics constructed from the same information set.