At the Inquire Europe Spring Seminar 2025 in Brussels, Dr. Theis Ingerslev Jensen, Assistant Professor of Finance at Yale School of Management, presented his joint research on the intersection of machine learning (ML), portfolio optimization, and trading costs. The work challenges a critical gap in asset pricing literature: the gap between predictive power and implementability.
The central insight is that while machine learning models are highly effective at forecasting stock returns, they often ignore frictions like transaction costs and market impact—factors that can erode theoretical profits and render optimized portfolios untradeable in practice. To bridge this gap, Jensen introduced the concept of the implementable efficient frontier: an out-of-sample, after-cost version of the classical Markowitz efficient frontier. This construct represents the best trade-offs between risk and return that an investor can realistically achieve, given trading frictions.
“Even though this portfolio is great before trading costs, the returns actually become negative once you account for them,” Jensen noted, referring to standard ML-based strategies.
To overcome this issue, he proposes Portfolio ML, a novel method that uses machine learning not just to predict returns, but to estimate portfolio weights directly. Crucially, Portfolio ML optimizes an investor’s actual utility function—one that explicitly accounts for trading costs—rather than relying on error minimization or return forecasting alone.
The approach addresses a critical issue: portfolio construction under frictions is inherently dynamic. Trading today impacts not just current performance, but future allocations and costs. Static, one-period optimization methods, while common in practice, fail to account for these intertemporal effects. Jensen highlighted this limitation, noting:
‘With trading cost, you shouldn’t just think about how your trades affect your portfolio today, but also how it affects the portfolio you can create tomorrow.’
Empirically, the model also reshapes our understanding of which asset characteristics matter most. In traditional, frictionless settings, signals like short-term reversal are often favored. But once trading costs are introduced, such signals lose relevance—they imply high turnover and excessive costs. Instead, the model reveals that value and quality characteristics are far more effective for investors facing trading constraints.
“Once you consider trading costs, it’s not about short-term alpha. You want persistent signals—and you want to trade slowly.”
The research is grounded in real-world data: U.S. equities, realistic price impact estimates, and various optimization benchmarks. Across all evaluation metrics—including net Sharpe ratio, utility, and implementation feasibility—Portfolio ML consistently delivers superior results. The methodology is made fully replicable, with the code being publicly available through GitHub.
By reframing how we evaluate performance and design strategies under realistic constraints, this work offers both a theoretical breakthrough and a practical toolkit for institutional investors navigating the friction-filled realities of modern markets.
View the research in full via : https://www.inquire-europe.org/event/joint-spring-seminar-2025-brussels/
The prize committee of Inquire Europe decided to award the prize for the best paper presented at the Joint Spring Seminar 2025 to the paper of Theis Ingerslev Jensen which he wrote with his co-authors Bryan Kelly (Yale School of Management), Semyon Malamud (Swiss Finance Institute), and Lasse Heje Pedersen (Copenhagen Business School). Click here to read more.