Technical Summary: How Many Bets Survive Stress? Retained Diversification in Asset Allocation

1. Problem Statement

The paper addresses the risk of diversification collapse: the phenomenon where a portfolio, constructed to appear well-diversified at the rebalance date based on current covariance estimates, loses its latent diversification when exposed to adverse dependence states during the holding period.

While traditional risk-based portfolio construction (e.g., Risk Parity, Minimum Variance) distributes variance across assets or latent covariance components at formation, the value of diversification depends on the dependence states realized after formation. In stress scenarios, correlations often rise, hedge relations (such as stock-bond hedges) weaken, and common factors absorb more portfolio risk. Consequently, assets that appear as distinct bets under the formation-date covariance matrix may become exposures to the same stress component, effectively reducing the number of independent risk directions. The paper argues that measuring diversification solely via the current covariance matrix overstates the durability of diversification and fails to capture the "diversification-collapse risk" inherent in multi-asset portfolios.

2. Methodology

Data and Scope

The study utilizes daily returns from 13 liquid ETFs spanning major asset classes (equities, nominal rates, credit, inflation-linked bonds, commodities, real estate, and the U.S. dollar) from April 12, 2007, to May 15, 2026. Portfolios are formed monthly, with validation tests covering the subsequent 63 trading days (a quarterly, unrebalanced holding period).

The Core Metric: Stress-Retained Effective Number of Bets (ENB)

The authors introduce Stress-Retained ENB, a measure of the effective diversification preserved across scenarios formed from information available at portfolio formation.

• Effective Number of Bets (ENB): Following Deguest, Martellini, and Meucci (2022), ENB is defined as the entropy-equivalent number of independent latent risk factors carrying portfolio variance. It is calculated as $ENB = \exp(-\sum p_j \ln p_j)$, where $p_j$ represents the variance share of the $j$-th principal component.
• Retention Mechanism: Unlike static ENB, retained ENB accounts for the evolution of the portfolio under a specific scenario law $Q_t$. It involves:
  1. Generating $h$-day return paths ($h=63$) based on formation-date information.
  2. Allowing portfolio weights to drift via buy-and-hold dynamics within each path.
  3. Estimating a path-specific covariance matrix $\Sigma_t^{(s)}$ for each scenario.
  4. Computing the average start-of-day ENB across the path using the drifted weights and the path-specific covariance.

Risk Metrics

To summarize the lower tail of the retained diversification distribution, the paper defines:

• ENB-at-Risk (EaR): The lower $\alpha$-quantile (baseline $\alpha=0.05$) of the retained ENB distribution. It answers how few effective bets remain in the worst 5% of scenarios.
• Conditional ENB Shortfall (CES): The conditional mean of retained ENB below the EaR threshold.
• Fragility: Defined as the difference between Current ENB and CES.

Scenario Construction

The baseline scenario law ($Q_t$) utilizes 300 prior-information return paths per rebalance date, constructed from:

• 40% block-bootstrap paths from the prior 252 trading days.
• 40% completed historical paths from prior "high-stress" months (selected via an expanding state index combining correlation, Treasury-loss pressure, and absorption ratios).
• 20% covariance-shock paths simulated from trailing means and eligible prior high-state covariance matrices.

Portfolio Construction and Validation

The study evaluates eight benchmark allocation rules (Equal Weight, 60/40, Inverse Volatility, Risk Parity, GMV, Maximum Diversification, FRP-MV, and MV-ENB6) within a constrained feasible set (long-only, ETF caps at 60%, asset-class caps).

• Validation: The realized path-mean ENB over the next 63 days is compared against ex-ante forecasts using check loss (for EaR) and identification moments (for CES).
• Calibration: The Probability Integral Transform (PIT) and Kolmogorov-Smirnov (KS) distances are used to assess how well the ex-ante scenario distributions match the realized retained ENB.

3. Key Results

Diversification Collapse and Stress

• Stress Impact: Stress-retained ENB declines significantly when broad financial-stress indexes (e.g., NFCI, ANFCI) indicate stress. High-stress states are associated with reduced effective bets, stronger cross-asset dependence, and weaker hedge relations.
• Mechanism: The decline in retained diversification is driven by increased average and downside correlations, a rise in stock-bond correlation (weakening the duration hedge), and higher absorption ratios (where the first few principal components explain more variance).

Portfolio Comparisons

• FRP-MV vs. GMV: The Factor-Risk-Parity Minimum-Variance (FRP-MV) portfolio achieves a higher Current ENB (10.58) than the Global Minimum-Variance (GMV) portfolio (7.81). However, in the stress tail, FRP-MV retains fewer effective bets (CES of 3.99 vs. 4.18 for GMV) and exhibits significantly higher fragility (6.59 vs. 3.63). This demonstrates that maximizing current ENB does not guarantee robustness in adverse states.
• Ranking: Portfolios with higher CES and EaR (indicating better retained diversification) consistently exhibit lower subsequent downside risk.

Predictive Power for Downside Risk

• Risk Reduction: Among matched feasible portfolios (controlled for current ENB, volatility, duration, and asset weights), those in the top tercile of CES or EaR experience significantly lower future losses. Specifically, high-retention portfolios show roughly 9% to 10% lower downside risk (measured by VaR loss, Expected Shortfall, downside volatility, and maximum drawdown) compared to low-retention portfolios.
• Forecast Accuracy:
  • EaR: The Stress EaR benchmark achieves a check loss of 0.1109, outperforming the Rolling Block-Bootstrap benchmark (0.1579) by 29.8%.
  • CES: Stress CES demonstrates the lowest mean absolute identification moment (0.9605), indicating superior calibration of the lower tail compared to point-covariance and rolling benchmarks.
  • Distributional Fit: The stress-scenario distribution tracks the full distribution of realized retained ENB more closely than current-covariance point estimates or rolling block-bootstrap distributions, with a KS distance of 0.070 compared to 0.134 for rolling bootstraps.

4. Significance and Claims

The paper claims to establish retained diversification as a viable ex-ante diagnostic for subsequent downside risk. Its primary contributions are:

1. Redefining Diversification Risk: It shifts the focus from static diversification (at formation) to diversification-collapse risk, quantifying how many effective bets survive adverse dependence states.
2. Superior Diagnostic Tools: It introduces EaR and CES as metrics that summarize the lower tail of retained diversification, providing a more accurate characterization of portfolio vulnerability than current-covariance ENB.
3. Decoupling Current and Retained ENB: The results show that portfolios can appear highly diversified at formation (high Current ENB) but possess low retained diversification in stress (low CES), as evidenced by the FRP-MV vs. GMV comparison.
4. Predictive Validity: The study provides empirical evidence that higher ex-ante retained ENB (measured by CES or EaR) predicts lower realized downside risk (VaR, ES, drawdown), even when controlling for standard risk factors.
5. Scenario Modeling: It demonstrates that stress-scenario distributions, constructed from prior information and high-state historical paths, offer better calibration and distributional fit for retained diversification than standard rolling window or point-covariance approaches.

The authors conclude that retained ENB measures the "fragility" of formation-date allocations, linking the loss of latent diversification in adverse states directly to the magnitude of subsequent portfolio losses.