Maximum Diversification Ratio

The MDR objective function maximizes the Diversification Ratio (DR), defined as:

DR = (wᵀσ) / √(wᵀΣw)

Where:

• w is the vector of portfolio weights
• σ is the vector of asset volatilities
• Σ is the covariance matrix

The numerator represents the weighted-average volatility if correlations were perfect. The denominator is the actual portfolio volatility. The ratio quantifies how much risk reduction is achieved through imperfect correlations. An MDR of 1.0 indicates no diversification benefit, while higher values signal greater diversification efficiency.

MDR portfolios exhibit extreme sensitivity to the correlation matrix, not just volatilities. The optimizer aggressively allocates capital to assets with low pairwise correlations to maximize the denominator reduction.

Key behavioral properties:

• Assets with low correlation to the rest of the portfolio receive higher weights, even if their standalone volatility is high
• Highly correlated asset clusters are penalized, often receiving near-zero allocations
• The resulting portfolio is the tangency portfolio when all assets have identical Sharpe ratios, linking MDR to mean-variance efficiency under specific assumptions

The MDR problem is formulated as a convex optimization with a fractional objective:

maximize (wᵀσ) / √(wᵀΣw) subject to ∑wᵢ = 1 and wᵢ ≥ 0

This can be transformed into a Second-Order Cone Program (SOCP) for efficient solving. The objective is quasi-concave, guaranteeing a unique global maximum. Unlike mean-variance optimization, MDR does not require expected return estimates—only the covariance matrix and volatilities. This eliminates the most error-prone input in portfolio construction, making MDR more robust to estimation error than mean-variance approaches.

MDR and Equal Risk Contribution (ERC) are distinct but related diversification frameworks:

• MDR maximizes the diversification ratio directly, often producing concentrated weights in low-correlation assets
• ERC equalizes each asset's marginal risk contribution, producing more balanced weight distributions
• MDR is equivalent to ERC only when all pairwise correlations are equal—a condition rarely met in practice
• MDR can be viewed as maximizing the Effective Number of Bets (ENB) under certain correlation structures

In practice, MDR portfolios tend to be more concentrated than ERC portfolios, particularly when a subset of assets exhibits exceptionally low correlations to the rest of the universe.

MDR portfolios face significant estimation error sensitivity in real-world applications:

• The optimizer exploits sampling errors in the correlation matrix, amplifying noise rather than signal
• Small changes in estimated correlations can cause large weight fluctuations between rebalancing periods
• Covariance shrinkage techniques (e.g., Ledoit-Wolf) are essential to stabilize MDR weights
• Constraining maximum asset weights or applying L1/L2 regularization to the covariance matrix improves out-of-sample performance
• The lookback window for covariance estimation critically impacts results—shorter windows increase responsiveness but amplify turnover

MDR portfolios theoretically capture the diversification return—the difference between the weighted-average geometric return of assets and the portfolio's geometric return:

Diversification Return ≈ (wᵀσ² - σ²_portfolio) / 2

This represents the rebalancing premium earned from systematically buying assets that have underperformed and selling those that have outperformed. MDR maximizes this premium by constructing the portfolio with the largest gap between average asset variance and portfolio variance. In practice, this premium is eroded by transaction costs, making the net benefit dependent on rebalancing frequency and market impact.