Machine Learning in Investing

What Really Works?

The promise of machine learning in investment management has captured significant attention in recent years. These methods attract interest due to their ability to exploit broader information sets and complex, nonlinear relations. Advanced algorithms claiming to identify profitable trading opportunities have proliferated across academic journals and investment pitches. But a crucial question remains: do these sophisticated techniques deliver real-world value, or do their benefits evaporate under practical constraints?

Mikheil Esakia and Felix Goltz, authors of the December 2025 study “What Drives the Performance of Machine Learning Factor Strategies?” tackled this critical question by examining what actually drives the performance of machine learning factor strategies when subjected to realistic implementation conditions.

What the Researchers Examined

Esakia and Goltz investigated two fundamental questions about machine learning approaches to stock selection:

1. Does the benefit come from using more information? Machine learning models can process far more variables when predicting stock returns.

2. Does the benefit come from capturing complex relationships? Neural networks can identify nonlinear patterns and interactions between variables that simpler linear models miss.

To disentangle these effects, the researchers created a 2×2 framework comparing:

• Linear vs. nonlinear models.

• Sparse (6 factors) vs. nonsparse (94 factors) predictor sets.

They then tested these models across three progressively realistic settings:

Standard Setting: The typical academic approach that sometimes includes hard-to-trade microcap stocks, uses factors that weren’t publicly known at the time, and ignores trading costs.

Intermediate Setting: Excludes microcaps and removes “factor hindsight” by only using publicly known predictors.

Realistic Setting: Additionally incorporates actual transaction costs into both portfolio construction and performance evaluation.

Their data set, covering the period June 1963-December 2021, consisted of common shares listed on the NYSE, AMEX, and NASDAQ and covered 94 firm-level characteristics. The findings reveal important limitations to machine learning’s effectiveness.

Key Findings

1. Context Matters Enormously

In the standard academic setting, both nonlinearity and expanded information sets appeared to add roughly equal value—about 1% per month each. However, in the intermediate setting, after eliminating hindsight and excluding microcaps, the incremental value of nonlinearity largely disappeared: return improvements from nonlinearity were marginally significant for nonsparse models (about 0.40% per month) and statistically insignificant for sparse models. Under realistic conditions, where costs shape both construction and evaluation, the overall return level declined from about 1.45% gross in the intermediate setting to about 0.79% net per month—a reduction of nearly 50%.

2. Information Beats Sophistication

Using more predictors proved more robust than using complex models. The value of an expanded information set persisted across all settings, though diminished in magnitude—monthly gains declined from about 1.20% to approximately 0.70%. In contrast, allowing for nonlinear relationships only helped when combined with a rich information set—and even then, primarily in the most unrealistic scenarios.

3. The “Important” Variables Change Dramatically

Features that appeared crucial in standard settings lost their predictive power under realistic constraints:

• Market capitalization: Added 48 basis points per month in the standard setting but actually reduced performance once microcaps were excluded.

• Short-term reversal: Contributed 41 basis points in the standard setting but only 3 basis points with transaction costs.

Meanwhile, other characteristics like share turnover volatility and momentum maintained consistent importance across settings.

4. Transaction Cost Management Is Critical

The researchers found that strategies designed to account for trading costs reduced costs by 66-72%—far more than their 25-40% reduction in turnover. This suggests these algorithms selectively avoided expensive trades rather than just reducing all trading activity.

Cost-aware strategies also dramatically reduced implementation challenges. The extreme “days-to-trade” metric (time needed to execute positions in illiquid stocks) fell by over 75% compared to cost-ignorant strategies.

5. Short-Selling Constraints Matter

When the analysis was restricted to long-only portfolios (no short-selling), the benefits of machine learning largely evaporated. In the realistic setting, only the combination of nonlinearity and expanded information delivered meaningful outperformance (27 basis points per month). Neither ingredient alone provided statistically significant value.

What the Researchers Concluded]

Esakia and Goltz concluded: “Combining nonlinearity with a broader information set still delivers meaningful gains in realistic settings, suggesting that the joint application of these ingredients to machine learning models remains valuable even under practical constraints.” They added: “Under short-sale constraints and in a realistic setting, the strategy based on ML signals still delivers meaningful outperformance over a market benchmark, but gains are markedly muted compared to the long-short case.”

Investor Takeaways