Optimal Versus Naive Diversification: How Inefficient Is the 1/N Portfolio Strategy?
Source: DeMiguel, V., Garlappi, L. & Uppal, R. (2009) · Review of Financial Studies 22(5), 1915–1953 · doi:10.1093/rfs/hhm075
Problem it solves
Mean-variance optimization (Markowitz, 1952) is optimal given the true inputs, but in practice expected returns and covariances are estimated with large error. The paper asks whether, under honest out-of-sample evaluation, optimized portfolios actually beat the trivial benchmark of putting 1/N of wealth in each of N assets.
TL;DR
Across fourteen sophisticated portfolio rules and seven empirical datasets, none consistently beats naive 1/N out of sample on Sharpe ratio, certainty-equivalent (CEQ) return, or turnover. Estimation error — especially in expected returns — overwhelms the theoretical gains from optimization. The estimation window required for sample-based mean-variance to beat 1/N is implausibly long: >3,000 months for 25 assets and >6,000 months for 50 assets, versus the ~120 months used in practice.
The method
Benchmark: equal weights w_i = 1/N, no estimation, no rebalancing optimization.
Competitors (14): sample mean-variance and minimum-variance; Bayes/shrinkage (Bayes-Stein, data-and-model, Bayesian diffuse-prior); the Ledoit-Wolf-type and Jagannathan-Ma constrained covariance rules; mean-variance with short-sale and other position constraints; combinations of these.
Evaluation: rolling out-of-sample, estimating on a 60- or 120-month window, scored by out-of-sample Sharpe ratio, CEQ return, and turnover (with statistical tests of Sharpe-ratio differences).
Assumptions & inputs
Monthly return data; investor with mean-variance preferences.
Inputs estimated from a rolling window — the analytics derive the window length needed for optimization to dominate 1/N as a function of N, the true Sharpe ratio of the optimal portfolio, and the Sharpe ratio of 1/N.
How to use it
Treat 1/N as a serious benchmark, not a strawman — beating it out of sample is genuinely hard.
When inputs are noisy, robust construction beats clever optimization: covariance shrinkage (Ledoit-Wolf), position/short-sale constraints (Jagannathan-Ma), and minimum-variance rules narrow the gap. The constrained minimum-variance (Jagannathan-Ma) policy performs best on Sharpe ratio but is still not statistically superior to 1/N in any of the seven datasets.
Treat estimation error as a first-class risk and prefer methods that degrade gracefully.
Limitations & pitfalls
Datasets are mostly portfolios (sectors, country/factor indices), where 1/N is already well-diversified; results can differ for individual-stock universes or when reliable return signals exist.
The verdict is about unconditional optimization with sample inputs; informative priors or factor structure can tilt the comparison.
Key references
DeMiguel, V., Garlappi, L. & Uppal, R. (2009) — Optimal Versus Naive Diversification — Review of Financial Studies
Markowitz, H. (1952) — Portfolio Selection — Journal of Finance
Ledoit, O. & Wolf, M. (2004) — Honey, I Shrunk the Sample Covariance Matrix — Journal of Portfolio Management
Jagannathan, R. & Ma, T. (2003) — Risk Reduction in Large Portfolios: Why Imposing the Wrong Constraints Helps — Journal of Finance
Provenance: verified/generated from the paper's full text.
