Diversification math

Definition

Diversification’s arithmetic: a portfolio’s variance is not the average of its components’ variances but a function dominated by their COVARIANCES — combining imperfectly correlated assets reduces risk without proportionally reducing expected return, the only structural free effect in portfolio construction (Markowitz’s founding insight). The math also states its own limits: systematic (common-factor) risk does not diversify away, and diversification’s benefit is a function of correlation, which is a regime variable.

How it works / structure

• The core arithmetic: for N equal-weight assets with average variance σ² and average pairwise covariance c, portfolio variance → c as N grows — idiosyncratic risk (σ² − c) washes out at roughly 1/N; the covariance floor remains. The floor IS the market/factor risk (strategy-factor-investing exposures).
• How many names: Statman (1987) and successors — most idiosyncratic variance reduction arrives by 20-30 uncorrelated names (later work argues more in modern data); the platform’s operational rule: count EFFECTIVE independent exposures (correlation-adjusted), not tickers (port-correlation-budgets — 30 tech stocks are ~3 exposures).
• The skewness case: Bessembinder (2018) — most individual stocks underperform T-bills over their lives; aggregate equity returns come from a small right tail — concentrated single-name portfolios face a NEGATIVELY SKEWED chance of holding the winners; diversification is partly a claim on the tail one cannot pre-identify.
• Correlation’s regime dependence: average correlations rise in stress (documented — the diversification benefit shrinks exactly when wanted, risk-correlation-exposure); cross-asset diversification inherits the stock-bond regime (regime-rate-environments).

When it applies

Book construction (effective-exposure counting as the standing diagnostic); strategy mixing (strategy-return correlations behave better than asset correlations when the strategies’ drivers genuinely differ — the platform’s multi-strategy case); concentration decisions (Bessembinder’s skew is the evidence AGAINST casual concentration and the honest frame for deliberate concentration theses).

Risk profile & failure modes

• Ticker-count theater: many names, one factor — the most common retail diversification failure (the effective- exposure diagnostic exists for this).
• Stress convergence: correlation matrices estimated in calm regimes overstate the benefit precisely at the tail (port-correlation-budgets stress matrices).
• Di-worsification: adding correlated mediocrity dilutes edge without reducing factor risk — diversification’s benefit applies to RISK; expected return is diluted linearly.
• Over-aggregation: strategy-level diversification claims need driver-level independence, not just historical correlation (two strategies long the same crowded factor measure independent until the unwind).

Evidence & limits

The variance arithmetic is mathematics. Statman (1987) and the N-question literature, and Bessembinder (2018) skewness evidence, are peer-reviewed. Correlation regime-dependence is documented (Longin-Solnik line of work cited in risk-correlation-exposure). The math promises variance reduction from imperfect correlation — nothing about returns, and nothing when correlations converge.

Falsifiable-thesis examples

Illustrations only, not signals:

• “This book’s effective independent exposures (correlation- adjusted) number at least 8 (diversification audit)” — falsified by the eigenvalue computation.
• “The 25-name equal-weight book will realize at least 30% lower vol than its average constituent this year (the arithmetic check)” — falsified by the realized pair.

Cross-references

• The framework layer: port-allocation-frameworks
• The matrix and its regimes: port-correlation-budgets, risk-correlation-exposure
• The factor floor: strategy-factor-investing
• The netting layer: port-exposure-netting