A new retrieval method sorts stocks by which ones are likely to move together tomorrow, not which ones looked alike yesterday.
Researchers describe Future-Aligned Soft Contrastive Learning (FASCL), a representation-learning framework for finding similar assets in a financial universe. Instead of grouping stocks by past price patterns or sector labels, FASCL trains on pairwise future return correlations as its supervision signal, using a soft contrastive loss that treats "similarity" as a continuous, not binary, target. The team also built a new evaluation protocol that checks whether retrieved assets actually share similar future trajectories, rather than just similar histories. Tested against 14 baseline methods on 5,631 US-listed securities, FASCL topped every rival on future return correlation and rank information coefficient at every retrieval depth and time horizon, led on trend consistency in 13 of 16 test cells, and produced the highest gross Sharpe ratio in a spread-trading backtest across every basket size.
The pitch matters because most asset-similarity tools quants actually use - sector classifications, historical correlation matrices - are backward-looking by design. They tell you what a stock did, not what it's about to do, and they're least reliable exactly when markets shift regimes and old relationships break down. A retrieval method trained explicitly to predict future co-movement targets that blind spot directly, which matters for anything built on finding "similar" assets: pairs trades, hedges, and spread strategies.
Still, a backtest is not a track record. FASCL's numbers come from one dataset of well-covered US tickers and a research paper's own evaluation protocol, not live trading. Plenty of promising quant methods have looked airtight in a backtest and then met a real order book.