Embedding similarity · Pair analysis · Mean reversionTraditional cross-sectional analysis relies on price correlation or sector classification to identify related companies. Our approach uses high-dimensional embeddings derived from comprehensive financial datasets to compute fundamental similarity. Two companies may operate in different sectors but describe nearly identical business models, risk factors, and competitive dynamics. Our embedding space captures this deep structural similarity in ways that standard classification systems cannot.
Once fundamentally similar pairs or baskets are identified, we monitor the valuation spread between them. When the spread exceeds historical norms — adjusted for fundamental changes detected in recent data — our systems flag it for investigation. The key analytical insight is that our similarity metric is continuously updated as new information arrives, allowing us to distinguish between temporary divergences and permanent fundamental differences.
Cross-sectional strategies require careful decomposition of what is driving observed spreads. We decompose every divergence into its signal loadings using our proprietary embedding dimensions and track exposure across all relevant factors. This granular decomposition ensures that we are identifying genuine analytical edges rather than inadvertent exposures to sectors, factors, or market regimes.