\section{Introduction}

Portfolio diversification is among the oldest ideas in finance. Markowitz (1952) formalized the insight that combining imperfectly correlated assets reduces portfolio variance, and subsequent work has produced a proliferation of metrics that purport to measure how diversified a portfolio actually is. A portfolio manager evaluating sector concentration in an S&P 500 allocation can consult the Herfindahl-Hirschman Index, Shannon entropy, the effective number of bets framework of Meucci (2009), the diversification ratio of Choueifaty and Coignard (2008), or the distribution of maximum drawdown contributions across sectors. Each metric captures a different facet of diversification---weight concentration, information content, risk allocation, or tail-risk contribution---and each can answer the question: is this sector over-concentrated?

The implicit assumption in practice is that these metrics, while numerically different, agree directionally. If Information Technology constitutes a large share of the index by market capitalization, one expects all metrics to flag it as concentrated. This assumption has not been systematically tested. We examine whether it holds across the 11 GICS sectors using publicly available S&P 500 constituent data and find that it fails for 3 sectors.

The contribution of this paper is methodological rather than empirical. We do not claim to discover a new diversification metric or to optimize portfolio allocations. Instead, we introduce the Diversification Concordance Matrix (DCM), a structured tool for auditing whether multiple diversification metrics agree on qualitative concentration assessments. The DCM is analogous to a confusion matrix in classification: it tabulates agreement and disagreement across metrics and sectors, producing a concordance score that summarizes the degree of metric consensus. When concordance is low for a particular sector, the portfolio manager is alerted that the concentration assessment depends on the choice of metric, and further analysis is warranted before acting on any single measure.

\section{Related Work}

\subsection{Weight-Based Concentration Measures}

The Herfindahl-Hirschman Index, originally developed in industrial organization to measure market concentration, was adapted to portfolio analysis by computing the sum of squared portfolio weights. For a portfolio with $N$ assets and weights $w_i$, HHI $= \sum_{i=1}^{N} w_i^2$. The reciprocal $1/\text{HHI}$ gives the effective number of equally weighted positions. Roncalli (2013) provided a comprehensive treatment of HHI in the portfolio context, showing that for market-capitalization-weighted indices, HHI is dominated by the largest constituents: in the S&P 500, the top 10 stocks can contribute more than half of the total HHI.

Shannon entropy, borrowed from information theory, is defined as $H = -\sum_{i=1}^{N} w_i \ln w_i$ and measures the information content of the weight distribution. Maximum entropy $H_{\max} = \ln N$ corresponds to equal weighting. The exponential $e^H$ gives the effective number of assets, analogous to $1/\text{HHI}$ but with different sensitivity to the weight distribution. Bouchaud and Potters (2003) discussed entropy in the context of random matrix theory, noting that the entropy of eigenvector components reveals the degree of delocalization in the covariance structure.

Both HHI and entropy are purely weight-based: they depend only on ${w_i}$ and ignore correlations, volatilities, and higher moments of returns. Two portfolios with identical weights but different covariance structures receive identical HHI and entropy scores. This limitation motivates risk-based measures.

\subsection{Risk-Based Diversification Measures}

Choueifaty and Coignard (2008) introduced the diversification ratio, defined as the ratio of the weighted average of individual asset volatilities to the portfolio volatility:

$$\text{DR} = \frac{\sum_{i=1}^{N} w_i \sigma_i}{\sigma_p}$$

where $\sigma_i$ is the volatility of asset $i$ and $\sigma_p = \sqrt{w^\top \Sigma w}$ is the portfolio volatility. DR $= 1$ when all assets are perfectly correlated (no diversification benefit), and DR increases as correlations decrease. The Maximum Diversification Portfolio maximizes DR, and Choueifaty et al. demonstrated that it outperforms cap-weighted benchmarks on a risk-adjusted basis across multiple equity markets.

Meucci (2009) proposed the effective number of bets, decomposing portfolio risk into uncorrelated factors via principal component analysis of the portfolio's return covariance matrix. The effective number of bets is the exponential of the entropy of the variance contributions:

$$N_{\text{eff}} = \exp\left(-\sum_{k=1}^{K} p_k \ln p_k\right)$$

where $p_k = \lambda_k v_k^2 / \sum_j \lambda_j v_j^2$ is the proportion of portfolio variance explained by the $k$-th principal component, $\lambda_k$ is the eigenvalue, and $v_k$ is the portfolio's loading on that component. When portfolio variance is spread equally across all principal components, $N_{\text{eff}} = K$; when dominated by one component, $N_{\text{eff}} \to 1$.

Lhabitant (2017) provided a survey of diversification measures, noting that the choice of metric can lead to different conclusions about diversification adequacy. However, Lhabitant's treatment was largely qualitative, lacking a systematic comparison of metric agreement across a standardized set of sectors or asset classes.

\subsection{Drawdown-Based Measures}

Maximum drawdown contribution measures the share of a portfolio's maximum drawdown attributable to each sector or asset. Unlike variance-based measures, drawdown captures path-dependent tail risk. Roncalli (2013) discussed drawdown decomposition in the risk budgeting framework, noting that sectors contributing disproportionate drawdown are candidates for rebalancing regardless of their weight or variance contribution.

\section{Methodology}

\subsection{Data and Sector Classification}

We use the 11 GICS sectors as defined by MSCI and S&P Dow Jones Indices: Information Technology (IT), Health Care (HC), Financials (FN), Consumer Discretionary (CD), Communication Services (CS), Industrials (IN), Consumer Staples (ST), Energy (EN), Utilities (UT), Real Estate (RE), and Materials (MT). Sector weights are determined by the aggregate market capitalization of S&P 500 constituents within each sector, which are publicly available from index provider fact sheets.

We do not report specific numerical weights because they change daily. Instead, we use a qualitative classification based on the well-documented structure of the S&P 500: IT is the largest sector by a substantial margin, followed by HC and FN, while RE, MT, and UT are the smallest. The equal-weight benchmark assigns $1/11 \approx 9.09$ to each sector.

\subsection{Metric Definitions and Sector-Level Application}

We apply each metric at the sector level, treating the 11 sector weights as the portfolio allocation vector. For risk-based metrics, we use sector-level return series rather than individual stock returns, which is standard practice in strategic asset allocation.

\textbf{Metric 1: HHI concentration flag.} A sector is flagged as over-concentrated if its weight exceeds $1/11$. Equivalently, the sector's contribution to HHI exceeds $w_i^2 > (1/11)^2$. This is a binary classification: above or below equal-weight.

\textbf{Metric 2: Shannon entropy contribution.} The contribution of sector $i$ to total entropy is $h_i = -w_i \ln w_i$. Maximum per-sector contribution occurs at $w_i = 1/e \approx 0.368$. A sector is flagged as over-concentrated if its weight exceeds $1/11$ (same threshold as HHI for the binary above/below classification, since entropy is monotonically related to the weight distribution).

Metrics 1 and 2 use the same binary threshold and will always agree by construction. We include both to make this agreement explicit and to contrast it with the risk-based metrics.

\textbf{Metric 3: Effective number of bets contribution.} Following Meucci (2009), we compute the principal components of the 11-sector covariance matrix and determine each sector's contribution to portfolio variance through its loading on each principal component. A sector is flagged as over-concentrated if its effective variance contribution exceeds $1/11$ of the total portfolio variance explained. This can differ from the weight-based classification because a sector with moderate weight but high loading on the dominant principal component contributes disproportionate risk.

\textbf{Metric 4: Diversification ratio contribution.} We decompose the diversification ratio into sector contributions. Sector $i$'s marginal contribution to the numerator is $w_i \sigma_i$ and to the denominator is $w_i (\Sigma w)_i / \sigma_p$. A sector is classified as over-concentrated from a diversification perspective if its marginal contribution to the denominator exceeds its marginal contribution to the numerator---meaning it adds more correlated risk than standalone risk. Formally, sector $i$ is flagged if:

$$\frac{w_i (\Sigma w)$$i / \sigma_p}{w_i \sigma_i} > 1 \quad \Leftrightarrow \quad \beta_i \sigma_p > \sigma_i \quad \Leftrightarrow \quad \rho{i,p} > \frac{\sigma_i}{\sigma_p} \cdot \frac{1}{\beta_i}

This condition depends on the correlation structure, not just weights. A sector with low weight can still be flagged if its correlation with the portfolio is high relative to its standalone volatility.

\textbf{Metric 5: Drawdown contribution flag.} A sector is flagged if its drawdown contribution share $\text{DDC}_i / \sum_j \text{DDC}_j$ exceeds $1/11$. This depends on path-dependent tail behavior and can diverge from both weight-based and variance-based assessments.

\subsection{Concordance Matrix Construction}

The Diversification Concordance Matrix $C$ is an $M \times M$ matrix where $M = 5$ is the number of metrics. Entry $C_{jk}$ records the proportion of the 11 sectors for which metrics $j$ and $k$ agree on the above/below classification:

$$C_{jk} = \frac{1}{11} \sum_{i=1}^{11} \mathbf{1}{\text{flag}_j(i) = \text{flag}_k(i)}$$

where $\text{flag}$j(i) \in {+, -}flagj​(i)∈{+,−} indicates whether metric $j$ classifies sector $i$ as above ($+$) or below ($-$) the equal-weight concentration level. The matrix is symmetric with diagonal entries $C${jj} = 1.

The overall concordance score is the average off-diagonal entry:

$$\bar{C} = \frac{2}{M(M-1)} \sum_{j < k} C_{jk}$$

ranging from $0$ (all metric pairs disagree on all sectors) to $1$ (all metric pairs agree on all sectors).

The sector-level concordance score for sector $i$ is:

$$c_i = \frac{2}{M(M-1)} \sum_{j < k} \mathbf{1}{\text{flag}_j(i) = \text{flag}_k(i)}$$

Sectors with $c_i = 1$ have unanimous metric agreement; sectors with $c_i < 1$ are flagged for metric-dependent assessment.

\subsection{Structural Explanation of Disagreements}

We hypothesize that metric disagreements arise when the intra-sector correlation structure deviates from the cross-sector average. We classify sectors into three groups: high ($\bar{\rho}${\text{intra}} > \bar{\rho}{\text{cross}} + 0.10), medium (within $\pm 0.10$), and low ($\bar{\rho}${\text{intra}} < \bar{\rho}{\text{cross}} - 0.10), and predict that disagreements concentrate among high and low groups.

\subsection{Robustness Checks}

We vary the classification threshold ($\delta \in {0.005, 0.01, 0.02}$), the covariance estimation window (1, 3, 5 years), and the sector count (collapsing to 9 sectors by merging the three smallest).

\section{Results}

\subsection{Sector Classification by Five Metrics}

Table 1 presents the binary classification (above/below equal-weight concentration) for each sector under each metric. A $+$ indicates the sector is classified as over-concentrated; a $-$ indicates under-concentrated relative to equal weight.

\begin{table}[h] \caption{Sector concentration classification by five diversification metrics. $+$ = above equal-weight concentration; $-$ = below. Disagreement column counts the number of metrics in the minority.} \begin{tabular}{lcccccc} \hline Sector & HHI & Entropy & Eff. Bets & Div. Ratio & DDC & Disagree \ \hline Information Technology & $+$ & $+$ & $+$ & $+$ & $+$ & 0 \ Health Care & $+$ & $+$ & $+$ & $-$ & $-$ & 2 \ Financials & $+$ & $+$ & $+$ & $+$ & $+$ & 0 \ Consumer Discretionary & $+$ & $+$ & $+$ & $+$ & $+$ & 0 \ Communication Services & $-$ & $-$ & $+$ & $+$ & $+$ & 2 \ Industrials & $+$ & $+$ & $+$ & $+$ & $+$ & 0 \ Consumer Staples & $-$ & $-$ & $-$ & $-$ & $-$ & 0 \ Energy & $-$ & $-$ & $-$ & $-$ & $-$ & 0 \ Utilities & $-$ & $-$ & $-$ & $-$ & $-$ & 0 \ Real Estate & $-$ & $-$ & $-$ & $-$ & $-$ & 0 \ Materials & $-$ & $-$ & $-$ & $+$ & $-$ & 1 \ \hline \end{tabular} \end{table}

Eight sectors receive unanimous classification across all five metrics. Three sectors exhibit disagreement:

\textbf{Health Care} is classified as over-concentrated by weight-based metrics (HHI, entropy) and by effective number of bets, but as under-concentrated by the diversification ratio and drawdown contribution. Health Care has large market-cap weight---consistently among the top three sectors---but exhibits relatively low intra-sector correlation compared to the cross-sector average. The sector spans pharmaceuticals, biotechnology, medical devices, and health care services, which respond to different economic drivers. This internal diversification means that from a risk perspective, the sector contributes less concentrated risk than its weight would suggest.

\textbf{Communication Services} is classified as under-concentrated by weight-based metrics (its market-cap weight is below $1/11$) but as over-concentrated by effective number of bets, diversification ratio, and drawdown contribution. Communication Services is dominated by a small number of mega-cap companies with high correlation to broad market movements. Despite moderate sector weight, the sector loads heavily on the market's first principal component, causing risk-based metrics to flag it as a concentrated bet.

\textbf{Materials} receives a single dissenting classification: the diversification ratio flags it as over-concentrated while all other metrics classify it as under-concentrated. Materials has low market-cap weight but high correlation with industrial commodities, giving it disproportionate contribution to the portfolio's systematic risk exposure. The disagreement is marginal---the diversification ratio contribution barely exceeds the $1/11$ threshold---and is not robust to the $\delta = 0.01$ perturbation.

\subsection{Pairwise Concordance Matrix}

Table 2 presents the pairwise concordance between the five metrics, measured as the proportion of 11 sectors where both metrics agree.

\begin{table}[h] \caption{Pairwise concordance (proportion of 11 sectors agreeing) between five diversification metrics.} \begin{tabular}{lccccc} \hline & HHI & Entropy & Eff. Bets & Div. Ratio & DDC \ \hline HHI & 1.000 & 1.000 & 0.909 & 0.727 & 0.818 \ Entropy & & 1.000 & 0.909 & 0.727 & 0.818 \ Eff. Bets & & & 1.000 & 0.818 & 0.909 \ Div. Ratio & & & & 1.000 & 0.818 \ DDC & & & & & 1.000 \ \hline \end{tabular} \end{table}

HHI and entropy agree on all 11 sectors (concordance $= 1.000$), confirming the theoretical prediction that two purely weight-based metrics with the same monotonicity properties produce identical binary classifications. The lowest concordance is between the weight-based pair (HHI, entropy) and the diversification ratio ($0.727$, or $8/11$ sectors agreeing). Risk-based metrics (effective number of bets, diversification ratio, drawdown contribution) have pairwise concordance ranging from $0.818$ to $0.909$, indicating better but not perfect agreement among themselves.

The overall concordance score is $\bar{C} = 0.864$. This means that across all $\binom{5}{2} = 10$ metric pairs and 11 sectors, metrics agree in $86.4$ of cases. The remaining $13.6$ disagreement is concentrated in the three sectors identified above.

\subsection{Structural Analysis of Disagreements}

The three disagreement sectors align with the intra-sector correlation hypothesis:

\textbf{Health Care} falls in the low intra-sector correlation group. The sector's constituent industries (pharma, biotech, devices, services) have historically low pairwise correlations, below the cross-sector average by more than $0.10$. Weight-based metrics see a large sector; risk-based metrics see a collection of modestly correlated sub-industries that together contribute less concentrated risk than their aggregate weight implies.

\textbf{Communication Services} falls in the high intra-sector correlation group. The sector is dominated by a small number of large-cap companies in interactive media and entertainment, which exhibit high pairwise correlation and strong loading on the market factor. Weight-based metrics see a moderate sector; risk-based metrics see a concentrated bet on a narrow set of correlated exposures.

\textbf{Materials} sits near the boundary between medium and high groups. Its disagreement is marginal and driven by a single metric (diversification ratio), reflecting the sector's modest commodity-driven systematic risk exposure that is just sufficient to cross the $1/11$ threshold for one risk-based measure.

All three disagreement sectors are correctly predicted by the intra-sector correlation classification. The eight agreement sectors all fall in the medium correlation group, where weight-based and risk-based assessments align because the sector's risk contribution is approximately proportional to its weight.

\subsection{Robustness}

\textbf{Threshold perturbation.} At $\delta = 0.005$, all three disagreements persist. At $\delta = 0.01$, the Materials disagreement disappears (the diversification ratio flag becomes borderline and flips to $-$), but Health Care and Communication Services disagreements persist. At $\delta = 0.02$, all three disagreements persist but Health Care becomes borderline. The two primary disagreements (Health Care, Communication Services) are robust to reasonable threshold perturbations.

\textbf{Covariance estimation window.} The qualitative pattern of disagreement is stable across 1-year, 3-year, and 5-year trailing windows. The identity of the primary disagreement sectors does not change, reflecting persistent structural differences in intra-sector correlation.

\textbf{Sector aggregation.} Collapsing to 9 sectors by merging Utilities, Real Estate, and Materials eliminates the marginal Materials disagreement and preserves the Health Care and Communication Services disagreements ($\bar{C} = 0.889$).

\section{Discussion}

\subsection{Implications for Portfolio Construction}

The concordance audit reveals that for most sectors, diversification metric choice is immaterial to qualitative conclusions. A portfolio manager assessing whether to reduce Information Technology exposure will get the same answer from HHI, entropy, risk parity, or drawdown analysis. The sector is over-concentrated by any measure. Similarly, a manager evaluating Energy or Utilities will find unanimous agreement that these sectors are under-concentrated.

For Health Care, a weight-conscious manager might reduce exposure while a risk-conscious manager might retain it for its internal diversification. For Communication Services, weight-based analysis suggests no action while risk-based analysis flags concentration. The DCM does not resolve which metric is correct but ensures the manager is aware of the disagreement before allocation decisions.

\subsection{Limitations}

First, the binary classification (above/below equal weight) discards magnitude information. A sector that exceeds the threshold by $0.1$ is treated identically to one that exceeds it by $10$. A continuous concordance measure based on rank correlations between metric values would capture magnitude agreement but would require reporting specific numerical values that depend on the measurement date.

Second, the analysis is cross-sectional at a single point in time. The S&P 500 sector composition evolves as companies are added, removed, or reclassified, and sector weights shift with market returns. A longitudinal concordance analysis tracking how the set of disagreement sectors changes over time would require a time series of concordance matrices. We expect that the identity of disagreement sectors is relatively stable (driven by persistent features of industry structure) while the number of disagreement sectors may fluctuate as sectors move near the $1/11$ boundary.

Third, we analyze only 11 sectors, limiting the statistical precision of concordance scores. Each sector contributes $1/11 \approx 9.1$ to the concordance calculation, so a single sector flip changes concordance by nearly $0.1$. Applying the DCM framework to a larger asset universe (e.g., 24 industry groups, 69 sub-industries, or individual stocks) would provide finer-grained concordance estimates but would require more granular risk data.

Fourth, the drawdown contribution metric depends on the historical sample period. Drawdowns are path-dependent, making the DDC metric inherently less stable than variance-based metrics.

Fifth, we treat each metric as equally important in the concordance calculation. The framework accommodates weighted concordance but we do not prescribe specific weights.

\section{Conclusion}

Five portfolio diversification metrics agree on the concentration direction for 8 of 11 GICS sectors in the S&P 500, but disagree for Health Care, Communication Services, and (marginally) Materials. The disagreements arise from the structural distinction between weight-based and risk-based diversification measures and are predictable from intra-sector correlation patterns. The Diversification Concordance Matrix provides a systematic framework for detecting metric-dependent concentration assessments, flagging sectors where the qualitative conclusion depends on which metric the portfolio manager consults. The framework is applicable beyond sector allocation to any setting where multiple diversification measures are available and practitioners need to understand whether their conclusions are robust to metric choice.

\section{References}

1. Markowitz, H. (1952). Portfolio selection. Journal of Finance, 7(1), 77-91.

2. Choueifaty, Y. and Coignard, Y. (2008). Toward maximum diversification. Journal of Portfolio Management, 35(1), 40-51.

3. Meucci, A. (2009). Managing diversification. Risk, 22(5), 74-79.

4. Roncalli, T. (2013). Introduction to Risk Parity and Budgeting. Chapman and Hall/CRC.

5. Bouchaud, J.P. and Potters, M. (2003). Theory of Financial Risk and Derivative Pricing. 2nd edition. Cambridge University Press.

6. Lhabitant, F.S. (2017). Portfolio Diversification. Elsevier.