The Limits of Similarity-Based Macro-Regime Forecasting

This paper examines whether similarity-based macro-regime forecasting, which conditions expected returns on the most similar historical macroeconomic states, retains predictive power as the forecasting object becomes less aggregated. We apply a fixed seven-variable macro-state representation, a Euclidean similarity ranking, and a consistent out-of-sample design to three return universes: the six Fama–French factor portfolios, the 49 Fama–French industry portfolios, and individual CRSP stocks over 1985 to 2024. The central finding is negative and economically material: at the individual-stock level, the regime model underperforms an expanding-window historical mean benchmark by approximately 13.40 percent in panel out-of-sample R², with a hit ratio barely above chance and gross ranking profits that are eliminated under realistic transaction costs. This failure contrasts sharply with the upper levels of the hierarchy. The framework reproduces the qualitative factor-level pattern of Mulliner et al. (2025), including a positive spread between the most and least similar state portfolios, while industry-level predictability is weak and selective, surviving only under the Clark and West adjustment and concentrating in cyclical sectors with high operating leverage. Cross-sectional tests show the model performs relatively better for larger and more actively traded stocks and worse for volatile stocks, yet observable characteristics explain less than one percent of the variation in stock-level performance, indicating that the deterioration is structural rather than sample-specific. The contribution is a boundary test of macro-regime similarity as a forecasting technology: the signal contains genuine information about systematic return components but is too coarse for individual-stock forecasting unless combined with firm-level information.