Empirical Estimator of Diversification Quotient

Basic Information

• Paper ID: 2506.20385
• Title: Empirical Estimator of Diversification Quotient
• Authors: Xia Han (Nankai University), Liyuan Lin (Monash University), Mengshi Zhao (Nankai University)
• Classification: q-fin.RM (Quantitative Finance - Risk Management)
• Publication Date: October 13, 2025
• Paper Link: https://arxiv.org/abs/2506.20385

Abstract

The Diversification Quotient (DQ), proposed by Han et al. (2025), is a novel portfolio diversification measurement method designed to quantify the risk reduction effect of diversification on portfolio risk level parameters. Based on a rigorous axiomatic framework, DQ effectively captures heavy-tailed distributions and common shocks while improving portfolio optimization efficiency. This paper investigates the convergence properties and asymptotic normality of empirical DQ estimators based on Value-at-Risk (VaR) and Expected Shortfall (ES), providing explicit formulas for asymptotic variance computation. Compared to the Diversification Ratio (DR) proposed by Tasche (2007), which may exhibit divergent asymptotic variance due to lack of location invariance, the DQ estimator demonstrates superior robustness across various distributional settings. The study validates asymptotic properties through simulations under elliptical distributions and constructs confidence intervals for DQ estimates on real financial data using AR-GARCH models and residual bootstrap methods.

Research Background and Motivation

Problem Background

1. Importance of Diversification Measurement: Diversification is a fundamental principle in portfolio management, aiming to reduce risk by combining assets with different risk characteristics. Accurately quantifying diversification effects is crucial for risk management and investment decision-making.
2. Limitations of Existing Methods:
   • Traditional diversification metrics such as Diversification Ratio (DR) and Diversification Benefit (DB) have significant deficiencies
   • Neglect the impact of heavy-tailed distributions and common shocks
   • Susceptible to manipulation through constant shifts or scaling operations
   • Difficult to apply in credit risk estimation
3. Theoretical Foundation of DQ: The DQ proposed by Han et al. (2025) is constructed based on six axioms (non-negativity, location invariance, scale invariance, rationality, normalization, and continuity), making it the first diversification metric based on a rigorous axiomatic framework.
4. Need for Statistical Inference: Despite DQ's theoretical advantages, the lack of comprehensive statistical property research limits its practical application in finance and economics.

Core Contributions

1. Established Consistency of Empirical DQ Estimators: Under mild continuity conditions, proved strong consistency of DQ estimators based on VaR and the class of convex risk measures (including ES).
2. Derived Asymptotic Normality and Variance Formulas: Under independent and identically distributed (i.i.d.) and α-mixing sequence conditions, established asymptotic normality of DQ^VaR and DQ^ES empirical estimators, providing explicit expressions for asymptotic variance.
3. Extended to Dependent Data: Extended asymptotic normality results to α-mixing sequences, ensuring reliability in complex dependent market environments with market shocks, volatility clustering, and serial correlation.
4. Comparative Analysis with DR: First established asymptotic normality of DR based on VaR and ES, revealing that DR may suffer from divergent asymptotic variance due to lack of location invariance.
5. Empirical Validation: Verified theoretical results through simulations under elliptical distributions and analysis of real financial data.

Methodology Details

Task Definition

Given a portfolio loss vector $X = (X_1, \ldots, X_n)$, based on a parametric risk measure class $\rho = (\rho_\alpha)_{\alpha \in I}$, the Diversification Quotient is defined as:

$DQ^\rho_\alpha(X) = \frac{\alpha^*}{\alpha}$

where $\alpha^* = \inf\{\beta \in I : \rho_\beta(\sum_{i=1}^n X_i) \leq \sum_{i=1}^n \rho_\alpha(X_i)\}$

Empirical Estimator Construction

VaR-Based DQ Estimator

Based on the alternative formula from Proposition 3: $DQ^{VaR}_\alpha(X) = \frac{1}{\alpha}P\left(S > \sum_{i=1}^n VaR_\alpha(X_i)\right)$

The empirical estimator is: $\hat{DQ}^{VaR}_\alpha(N) = \frac{1}{N\alpha}\sum_{k=1}^N \mathbf{1}_{\{\sum_{i=1}^n X_i^{(k)} > \sum_{i=1}^n \hat{x}^{VaR_\alpha}_i\}}$

ES-Based DQ Estimator

$\hat{DQ}^{ES}_\alpha(N) = \frac{1}{N\alpha}\min_{r \in (0,\infty)} \sum_{k=1}^N \left[\left(r\sum_{i=1}^n (X_i^{(k)} - \hat{x}^{ES_\alpha}_i) + 1\right)_+\right]$

Theoretical Analysis Framework

Consistency Analysis

Theorem 1: Under the conditions of law-invariant convex risk measure classes or VaR classes, if $\beta \mapsto \rho_\beta(S)$ is a strictly decreasing function, then the empirical DQ estimator converges strongly to the true value.

Asymptotic Normality

Theorem 2 (Asymptotic Normality of DQ^VaR): Under regularity conditions, $\sqrt{N}(\hat{DQ}^{VaR}_\alpha(N) - DQ^{VaR}_\alpha(X)) \xrightarrow{d} N(0, \sigma^2_{VaR})$

where $\sigma^2_{VaR} = A^T_{VaR}\Sigma_{VaR}A_{VaR}$, $A_{VaR} = (\frac{g(t_{n+1})}{\alpha f_1(t_1)}, \ldots, \frac{g(t_{n+1})}{\alpha f_n(t_n)}, -\frac{1}{\alpha})$

Theorem 3 (Asymptotic Normality of DQ^ES): Similarly established asymptotic normality of the ES-based DQ estimator.

Technical Innovations

1. Location Invariance Advantage: Unlike DR, DQ possesses location invariance, i.e., $DQ^\rho_\alpha(X+c) = DQ^\rho_\alpha(X)$, avoiding variance explosion problems when centering portfolios.
2. Extension to Dependent Data: Through α-mixing sequence theory, extended results to more realistic financial time series settings.
3. Nonparametric Approach: Employs empirical distribution methods, avoiding risks of distributional misspecification and naturally capturing fat tails and asymmetry in financial returns.

Experimental Setup

Simulation Study

• Elliptical Distribution Framework: Considers multivariate normal distribution $X \sim N(\mu, \Sigma)$ and multivariate t-distribution $Y \sim t(\nu, \mu, \Sigma)$
• Equicorrelation Model: $\Sigma = (\sigma_{ij})_{n \times n}$, where $\sigma_{ii} = 1$, $\sigma_{ij} = r$ (when $i \neq j$)
• Parameter Settings: $r = 0.3$, $n = 5$, $\nu = 3$, $\alpha = 0.1$, sample size $N = 5000$, 2000 replications

Empirical Data Analysis

• Data Source: Daily data of the 5 largest market-cap stocks in the S&P 500 (2010-2025)
• Stock Selection: XOM (Energy), MSFT (Information Technology), BRK/B (Finance), WMT (Consumer Staples), JNJ (Healthcare)
• Modeling Method: AR(1)-GARCH(1,1) model with Student-t distributed innovations
• Bootstrap Method: Residual bootstrap to construct 95% confidence intervals, rolling window of 500 trading days

Evaluation Metrics

• Accuracy of asymptotic variance
• Coverage rate of confidence intervals
• Sensitivity analysis under different parameters

Experimental Results

Simulation Results Validation

Asymptotic Normality Verification

• Normal Distribution Case: DQ^VaR_α estimates match well with $N(0.27, 1.88/N)$
• t-Distribution Case: DQ^VaR_α estimates closely align with $N(0.45, 2.52/N)$
• ES Estimator: Similarly verified asymptotic normality

Parameter Sensitivity Analysis

1. Effect of Confidence Level α: Asymptotic variance decreases with increasing α, consistent with theoretical expectations
2. Effect of Correlation Coefficient r: Variance first increases then decreases, reflecting combined effects of diversification and correlation
3. Effect of Number of Assets n: Variance monotonically decreases in Gaussian model; decreases after n>5 in t-distribution
4. Effect of Degrees of Freedom ν: Variance decreases with increasing ν as heavy-tail effects weaken

Empirical Data Results

Confidence Interval Construction

• DQ^VaR: Variance of 0.0096 at α=0.05, 0.0066 at α=0.1
• DQ^ES: Variance of 0.0089 at α=0.05, 0.0060 at α=0.1
• COVID-19 Period: Confidence intervals noticeably widen, reflecting increased market volatility

Comparison with DR

• DQ estimator's asymptotic variance remains bounded across various settings
• DR may exhibit variance explosion when portfolios are centered
• Validates DQ's location invariance advantage

Key Findings

1. Robustness: DQ estimators perform stably across different distributional settings
2. Practicality: Bootstrap confidence intervals provide reliable uncertainty quantification for practical applications
3. Superiority: Compared to DR, DQ avoids location sensitivity issues

Related Work

Diversification Measurement Research

• Traditional Methods: Tasche (2007)'s Diversification Ratio, Embrechts et al. (2009)'s Diversification Benefit
• Limitations: Lack axiomatic foundation, susceptible to manipulation, neglect extreme events

Statistical Inference for Risk Measures

• VaR Estimation: Bahadur (1966)'s empirical VaR consistency and asymptotic normality
• ES Estimation: Scaillet (2004)'s nonparametric kernel estimation, Chen (2008)'s extension to dependent data
• Related Developments: Extensions by Asimit et al. (2019), Bellini et al. (2022) to broader risk measure classes

Positioning of This Paper

Fills the gap in statistical inference theory for DQ, providing a solid statistical foundation for this emerging diversification measure.

Conclusions and Discussion

Main Conclusions

1. Established a complete statistical theory framework for empirical DQ estimators
2. Proved DQ's robustness advantages compared to traditional DR
3. Provided theoretical support for DQ's application in practical financial risk management

Limitations

1. Regularity Conditions: Requires existence and continuity of density functions
2. Computational Complexity: ES-based DQ estimation involves optimization problems
3. Sample Size Requirements: Tail risk estimation requires sufficient extreme event data

Future Directions

1. Portfolio Optimization: Incorporate DQ as constraints in optimization frameworks
2. Axiomatic Extension: Develop more general axiomatic theory independent of preset risk measures
3. High-Frequency Data Application: Extend to high-frequency financial data and real-time risk monitoring

In-Depth Evaluation

Strengths

1. Theoretical Rigor: Provides a complete statistical inference framework from consistency to asymptotic distribution
2. Practical Value: Bootstrap confidence intervals provide actionable tools for practical applications
3. Innovation: First to establish asymptotic theory for DQ and DR, revealing important robustness differences
4. Comprehensiveness: Covers both independent and dependent data, balancing theory and empirics

Weaknesses

1. Technical Threshold: Requires strong background in probability theory and stochastic processes
2. Computational Burden: Some estimators (e.g., DQ^ES) have high computational complexity
3. Assumption Limitations: Elliptical distribution assumptions may not apply to all financial data

Impact

1. Academic Value: Makes important contributions to diversification measurement theory
2. Practical Significance: Provides new tools for financial institutions' risk management
3. Reproducibility: Provides detailed algorithms and implementation details

Applicable Scenarios

• Risk management in banks and insurance companies
• Portfolio optimization and asset allocation
• Systemic risk assessment by regulatory authorities
• Academic research on diversification effects

References

This paper is primarily based on the following important literature:

• Han, X., Lin, L., Wang, R. (2025). Diversification quotients: Quantifying diversification via risk measures. Management Science.
• Tasche, D. (2007). Capital allocation to business units and sub-portfolios: The Euler principle.
• Bahadur, R. R. (1966). A note on quantiles in large samples.
• Chen, S. X. (2008). Nonparametric estimation of Expected Shortfall.