Previous studies on random-forest–based portfolio construction show that fundamental variables can be organized into decision-tree leaves that behave like small, rule-based “micro-strategies,” and that combining the most profitable leaves can beat a broad market index. Yet these results usually assume highly idealized portfolios: they hold a large number of very small positions, ignore transaction costs, and allow full rebalancing every period. Such designs are hard to trade in real settings. This paper offers a practical variant. We keep the predictive core—a random forest trained on fundamental data—but change only the portfolio-formation layer to (i) cap the number of holdings through a cardinality constraint, (ii) enforce a minimum position size so that all weights are economically meaningful, and (iii) measure returns net of proportional transaction costs. Using S&P 500–like historical data, we find that these implementation-aware portfolios preserve most of the excess return delivered by the original unconstrained approach, while achieving materially lower turnover and a cleaner, more realistic weight distribution. This indicates that tree-based stock selection can be adapted to real-world trading frictions with only modest performance sacrifice.
Previous studies on random-forest–based portfolio construction show that fundamental variables can be organized into decision-tree leaves that behave like small, rule-based “micro-strategies,” and that combining the most profitable leaves can beat a broad market index. Yet these results usually assume highly idealized portfolios: they hold a large number of very small positions, ignore transaction costs, and allow full rebalancing every period. Such designs are hard to trade in real settings. This paper offers a practical variant. We keep the predictive core—a random forest trained on fundamental data—but change only the portfolio-formation layer to (i) cap the number of holdings through a cardinality constraint, (ii) enforce a minimum position size so that all weights are economically meaningful, and (iii) measure returns net of proportional transaction costs. Using S&P 500–like historical data, we find that these implementation-aware portfolios preserve most of the excess return delivered by the original unconstrained approach, while achieving materially lower turnover and a cleaner, more realistic weight distribution. This indicates that tree-based stock selection can be adapted to real-world trading frictions with only modest performance sacrifice.
