For decades, traditional portfolio theory has centered around maximizing expected return relative to risk—often summarized by the Sharpe ratio. But for real-world investors, particularly those managing large pools of capital, this framework omits a crucial reality: transaction costs. These costs not only eat into returns but fundamentally alter the shape of the opportunity set itself. Ignoring them can lead to portfolios that are elegant in theory but costly or even unprofitable in practice.
The Real Cost of Overlooking Frictions
Most machine learning applications in finance follow a two-step approach: predict returns using signals, then optimize portfolios based on those predictions. While intuitive, this process often fails in live trading because it doesn’t account for the costs associated with implementing the resulting strategies. Frequent rebalancing, heavy exposure to illiquid securities, and excessive turnover all generate significant friction. The result? Attractive backtests that crumble in real markets.
Ignoring trading costs pushes portfolio strategies toward high-turnover, short-horizon signals like one-month reversals. These signals may show promise in isolated datasets but are typically costly to exploit. This is especially problematic when strategies are scaled up, as even moderate trade sizes can move markets and amplify costs.
The Implementable Efficient Frontier
The traditional efficient frontier depicts a clean, upward-sloping relationship between risk and expected return. But this is a theoretical construct that assumes frictionless trading. Once transaction costs are introduced, the frontier bends sharply downward. Increased portfolio risk implies larger positions and higher trading volumes, which escalate costs and reduce net returns.
For large asset managers, this bend isn’t just theoretical—it’s operational. As portfolios grow, market impact becomes a real constraint. A $10 billion portfolio can’t rapidly shift capital without influencing prices. The “implementable efficient frontier” reflects this reality. It shows how net Sharpe ratios decline as risk increases, limiting the benefits of leverage and risk-taking for large players.
Why Portfolio Size Changes the Game
Smaller investors can afford to chase high-turnover strategies because their trades are less likely to move prices. But large institutions must think differently. At scale, portfolio choices are shaped not just by expected return and volatility, but by liquidity, capacity, and cost efficiency.
Empirical evidence shows that as assets under management grow, optimal portfolios shift away from illiquid positions and signals that decay quickly. Larger investors are naturally pushed toward more stable, longer-horizon predictors and liquid securities that can absorb substantial flows without significant slippage.
Rethinking the Learning Process
Rather than treating transaction costs as an afterthought, a more effective approach is to build them directly into the optimization process. This is where the concept of “portfolio machine learning” comes in. Instead of first predicting returns and then constructing portfolios, this method trains algorithms to learn optimal portfolio weights directly—factoring in expected returns, volatility, and trading costs in a single step.
This integrated approach ensures that the model prioritizes realistic and sustainable signals. It steers away from transient patterns in illiquid stocks and toward features with lasting predictive power. It also enables the model to balance alpha potential against holding costs, which is especially important when positions must be held over longer horizons due to frictions.
Practical Implications and Outcomes
The research shows that when transaction costs are included in the learning process, the feature importance of signals shifts significantly. While traditional models overemphasize short-term reversals, the cost-aware model favors value and quality—signals associated with more liquid and stable opportunities.
Importantly, the portfolio machine learning method delivers outperformance. In out-of-sample tests, it improves Sharpe ratios by 20% and utility by 60% compared to traditional methods. These gains aren’t just statistical—they stem from a more realistic view of what can be implemented profitably in actual markets.
Final Thoughts
For investors seeking to build scalable, resilient strategies, transaction costs can no longer be an afterthought. They are a central component of portfolio construction. Whether using traditional methods or cutting-edge machine learning, integrating these frictions into the design phase is essential.
The implementable efficient frontier reminds us that not all alpha is accessible, and not all risk is worth taking. The future of portfolio optimization lies in approaches that account for the real cost of execution—and learn to navigate around it.