Capturing the many, varied relationships between financial asset price movements is essential for several financial tasks. Conventional methods, which rely on pairwise similarity measurements, fail to capture the subtlety and nuance that define these relationships. Inspired by this need for more sophisticated approaches to modeling, we outline a task-agnostic, self-supervised framework to learn embedding representations for financial assets, by using only their historical returns. Rather than computing similarity directly on raw asset returns, this approach encodes common co-occurrence of similar return fluctuations in a latent embedding space. Using clustering, nearest neighbor, and sector classification experiments, we demonstrate that the learned embeddings are task-agnostic. We then propose new context selection strategies aimed at the downstream task of portfolio optimization and minimizing volatility. We evaluate our approach in a long-only hedging experiment on more than 5 years of out-of-sample data. A comparative analysis with traditional benchmarks reveals statistically significant performance benefits accruing to our embedding-based approach.