• Portfolio construction is reframed as an ensemble learning problem rather than pure optimisation
• Diversification is designed directly rather than inferred from noisy inputs
• Holding many assets does not guarantee diversified decision making
• Lack of predictive diversity explains why portfolios fail under stress
• Small sacrifices in forecast accuracy can improve robustness and Sharpe ratios
• Diversity can be introduced at both learning and asset selection stages
• Adaptation is possible but must be governed with clear limits and controls
• Rising model complexity increases the risk of synchronized failure
• Risk oversight should focus on model ecosystems, not individual models
• Monitoring behavior matters more than monitoring outcomes

The following article explores a paper by Alejandro Rodriguez Dominguez, Head of Quantitative Analysis and Artificial Intelligence at MiraltaBank, and co-authors colleagues,  Muhammad Shahzad and Xia Hong that proposes a new way of thinking about portfolio construction and diversification.

Instead of treating diversification as a by-product of optimisation and estimated inputs, the work reframes portfolio allocation as a multi-hypothesis prediction problem, where each asset is treated as a hypothesis and the portfolio itself is modelled as a structured ensemble.

The key idea is to establish a direct, formal link between diversity in predictive models and out-of-sample portfolio risk diversification, and to show how this diversity can be controlled parametrically before optimisation, both at the learning stage and during asset selection.

The framework is designed to be compatible with existing portfolio construction methods and motivated by practical risk management concerns such as robustness, generalisation, and behaviour under stress.

The following questions explore the implications of this approach from both a quantitative and a risk-practitioner perspective.

You frame portfolio construction as a supervised ensemble learning problem rather than a traditional optimisation exercise. What practical advantages does this perspective bring for understanding and managing portfolio risk?

Traditionally, portfolio construction is framed as an optimisation problem where diversification is an indirect outcome of estimated inputs - expected returns, covariances, or scenarios.

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.

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

Each asset is treated as a hypothesis, the portfolio becomes the ensemble, and the portfolio return is the aggregation of multiple predictive views. Under squared loss, the theory tells us that the optimal ensemble combiner is the arithmetic mean, which naturally aligns the learning target with the equal-weighted portfolio.

The practical advantage for risk management is control. Instead of relying on the covariance matrix to imply diversification, we can design diversification directly into the learning and decision process, before any optimisation takes place.

This allows risk teams to treat diversification as a tunable parameter, rather than an assumption that may or may not hold once the model is deployed.

In short, this perspective shifts diversification from something we estimate to something we engineer, which is much easier to govern, monitor, and explain.

Your work links ensemble diversity directly to risk diversification using information-theoretic and geometric ideas. How does this help explain why some portfolios that look diversified on paper still fail out of sample?

This is one of the central motivations behind the work. Many portfolios look diversified because they hold many assets, sectors, or factors, but they are not diversified in how decisions are made.

When markets move into stress regimes, correlations spike and those portfolios suddenly behave as if they were one concentrated bet.

The ensemble perspective helps explain this because it focuses on diversity of predictions, not just diversity of holdings.

Through the bias–variance–diversity decomposition, diversity appears explicitly as a term that improves generalisation. If predictors are too similar, the ensemble overfits, and the portfolio inherits that fragility.

What the paper shows empirically is that portfolios built from more diverse predictors or asset sets remain better diversified out of sample, even when individual forecasts are imperfect.

This explains why portfolios that appear diversified ex-ante can still fail: they lack diversity in their underlying decision pathways.

From a risk point of view, this reframes the question from “Are my assets uncorrelated?” to “Do my models and signals fail differently under stress?” That distinction is critical.

 The Quality–Diversity trade-off is a central theme in your session. How should practitioners think about balancing return prediction accuracy against the need for robust diversification?

This trade-off is very familiar to risk practitioners, even if it’s not always framed this way. Highly accurate forecasts often rely on fragile structure, while more diverse signals may look noisier but tend to fail less catastrophically.

The key insight from the work is that diversification can be introduced before optimisation, in two complementary ways: during the learning of individual predictors and during asset selection. Both are controlled parametrically.

This means practitioners can deliberately trade a small amount of forecast accuracy for greater robustness and stability.

One of the more counterintuitive findings is that portfolios can improve out-of-sample Sharpe ratios even when average predicted returns decline.

This happens because diversification reduces downside risk and improves generalisation. From a risk management perspective, this is exactly the trade-off we want to manage consciously rather than leaving it to the optimiser.

So the practical message is: don’t optimise accuracy in isolation. Optimise the behaviour of the system under uncertainty.

Looking ahead, do you see ensemble-based portfolio frameworks becoming more adaptive in real time, and what challenges would need to be solved to make that viable in production environments?

Yes, I do see these frameworks becoming more adaptive, but adaptability needs to be treated as a risk-controlled feature, not just a performance enhancement.

The framework naturally allows adaptation because the diversity parameters can be adjusted as market conditions change.

However, the experiments show that optimal diversity levels differ across regimes, and in crisis periods the relationship becomes more complex, sometimes with multiple optimal ranges.

The main challenges are therefore not computational, but organisational and governance-related. You need:

• robust regime monitoring,
• limits on how fast and how far diversity parameters can move,
• and clear validation and escalation rules. 

In production, diversity parameters should be governed much like exposure limits or leverage constraints. If those guardrails are in place, ensemble-based frameworks offer a structured way to adapt portfolios without introducing instability.

As model complexity and AI usage increase in portfolio management, how do you expect the diversity–capacity trade-off to influence future approaches to model selection and risk governance?

As models become more complex, the biggest risk is no longer simple overfitting, but synchronised failure. High-capacity models tend to learn similar structure unless diversity is explicitly enforced, which leads to hidden concentration risk.

This shifts risk governance away from evaluating individual models and toward evaluating model ecosystems. In that context, diversity becomes a measurable and controllable variable, not a qualitative aspiration.

I expect future risk frameworks to monitor diversity in much the same way they monitor concentration or exposure today: something that is designed, stress-tested, and reviewed regularly.

As AI becomes more prevalent, diversity will be essential not just for performance, but for resilience.

What are the main failure modes of this framework, and how would a risk team detect them early?

The main failure mode is diversity collapse, where predictors or selected assets become too similar, often due to regime shifts or data leakage.

Another risk is excessive diversity, where signal quality degrades to the point where the portfolio becomes noisy rather than robust.

Early detection comes from monitoring ensemble-level metrics, not just portfolio performance. These include measures of prediction similarity, changes in ensemble variance, and sudden increases in turnover or drawdown sensitivity.

The key is to monitor behaviour, not just outcomes. By doing that, risk teams can intervene before performance deteriorates materially.

 How would you explain and defend this approach in front of a risk committee or regulator?

I would frame it as a risk-first approach rather than a performance-first one. The core idea is that diversification is being explicitly designed and controlled, rather than assumed through statistical estimates.

The parameters that control diversity are transparent, interpretable, and bounded. They can be documented, stress-tested, and reviewed just like any other risk control. Importantly, the framework does not remove traditional risk measures - it complements them.

That makes it easier to explain, audit, and justify than many black-box optimisation approaches.

How does this framework relate to existing robust or risk-parity-style portfolio methods?

Most robust or risk-parity approaches focus on stabilising the optimisation step - shrinking covariances, reweighting risk, or enforcing equal contributions. This framework operates earlier in the pipeline, by stabilising the decision inputs themselves through diversity.

They are complementary rather than competing approaches. In fact, the paper shows that introducing diversity at the asset selection stage improves the performance of many existing methods.

So rather than replacing established risk techniques, this framework provides an additional layer of protection.

 Does this work change how we should define diversification as a risk concept?

Yes, I think it broadens the definition. Diversification is not just about holding many assets or reducing correlations; it’s about ensuring independent sources of risk and independent ways of being wrong.

In AI-driven portfolio construction, that distinction becomes critical. Structural diversification without behavioural diversification is fragile. This work formalises that idea and gives practitioners tools to act on it.

In conclusion, ultimately, this work reframes diversification from something we observe after optimisation to something we design before capital is deployed.

That shift matters, because designed risk controls are easier to govern, stress-test, and trust - especially in environments where models and data can fail together.”