Stock Embeddings: Representation Learning for Financial Time Series †

Abstract

Identifying meaningful and actionable relationships between the price movements of financial assets is a challenging but important problem for many financial tasks, from portfolio optimization to sector classification. However, recent machine learning research often focuses on price forecasting, neglecting the understanding and modelling of asset relationships. To address this, we propose a neural model for training stock embeddings that harnesses the dynamics of historical returns data to reveal the nuanced correlations between financial assets. We describe our approach in detail and discuss several practical ways it can be used in the financial domain. Specifically, we present evaluation results to demonstrate the utility of this approach, compared to several benchmarks, in both portfolio optimization and industry classification.