Skip to content

Content Library

Back to Articles

Article April 3, 2026

Rethinking Diversification for a Model Driven Investment World

Rethinking Diversification for a Model Driven Investment World

Alejandro Rodriguez Dominguez, Head of Quantitative Analysis and Artificial Intelligence, MiraltabankLinkedIn profile for Alejandro Rodriguez Dominguez

As AI reshapes portfolio construction, traditional diversification is proving fragile. Alejandro Rodriguez Dominguez argues that embedding diversity directly into decision-making frameworks can improve resilience, governance, and out-of-sample performance
  • Traditional optimisation often produces fragile diversification that fails under stress
  • Ensemble learning reframes diversification as something designed, not assumed
  • True diversification depends on diversity of decisions, not just asset holdings
  • Quality–diversity trade-off allows better control of robustness versus accuracy
  • Portfolios can improve out-of-sample performance even with lower predicted returns
  • AI models risk synchronized failure without enforced diversity
  • Diversity should be treated as a measurable governance parameter
  • Key risks include diversity collapse and excessive noise from over-diversification 

As portfolio construction becomes increasingly driven by models and artificial intelligence, traditional assumptions about diversification are being challenged.

What once appeared to be a reliable outcome of optimization is now being exposed as fragile under stress, prompting risk and investment teams to rethink how resilience is built into portfolios.

According to Alejandro Rodriguez Dominguez, Head of Quantitative Analysis and Artificial Intelligence at Miraltabank, the problem lies in how diversification has historically been treated.

“Traditionally, portfolio construction is framed as an optimisation problem where diversification is an indirect outcome of estimated inputs,” he explained.

“From a risk management perspective, that’s problematic because those inputs are noisy and unstable, and diversification often looks good in sample but disappears out of sample.”

Instead, Rodriguez Dominguez advocates reframing portfolio construction as an ensemble learning problem.

In this framework, each asset represents a hypothesis, while the portfolio itself becomes the aggregation of multiple predictive views.

“By reframing portfolio construction as a supervised ensemble learning problem, we make diversification explicit rather than incidental,” he said.

This shift has significant implications for risk management. Rather than relying on correlations and covariance matrices to imply diversification, practitioners can design it directly into the decision-making process.

“This perspective shifts diversification from something we estimate to something we engineer,” he noted. “That makes it much easier to govern, monitor, and explain.”

The distinction becomes particularly important during periods of market stress. Many portfolios that appear diversified on paper fail precisely when diversification is needed most.

Rodriguez Dominguez attributes this to a lack of diversity in decision pathways rather than asset holdings.

“Many portfolios look diversified because they hold many assets, sectors, or factors, but they are not diversified in how decisions are made,” he said.

When markets shift into stressed regimes, correlations tend to converge, and portfolios behave as if they were concentrated exposures. The ensemble framework offers a different lens.

“The question becomes not ‘Are my assets uncorrelated?’ but ‘Do my models and signals fail differently under stress?’” he explained. “That distinction is critical.”

Central to this approach is what Rodriguez Dominguez describes as the quality–diversity trade-off.

Highly accurate models often rely on fragile structures that may break down under changing conditions, while more diverse signals can provide greater stability even if they appear less precise.

“Practitioners can deliberately trade a small amount of forecast accuracy for greater robustness and stability,” he said.

This can lead to counterintuitive outcomes. In some cases, portfolios can improve their out-of-sample performance even as predicted returns decline. “Portfolios can improve out-of-sample Sharpe ratios even when average predicted returns decline,” Rodriguez Dominguez explained. “This happens because diversification reduces downside risk and improves generalisation.”

As AI adoption accelerates, this trade-off is becoming more important. High-capacity models tend to learn similar patterns, increasing the risk of synchronized failure across systems.

“As models become more complex, the biggest risk is no longer simple overfitting, but synchronised failure,” he warned. “High-capacity models tend to learn similar structure unless diversity is explicitly enforced.”

Events you may be interested in

View all of our upcoming events

Sign Up to Our Newsletter

Sign up

Connect Plus

Unlock more with Connect Plus membership

This content is available to Connect Plus members.