Just Split It Evenly: The Paper That Showed 1/N Beats the Fancy Models

Imagine you have spent decades building increasingly clever portfolio construction models. Then someone runs a fair race between your models and a rule a child could follow: take your money, divide it by the number of assets, and put an equal slice in each one. That is the 1/N rule, and in 2009 Victor DeMiguel, Lorenzo Garlappi and Raman Uppal published the race results. The clever models did not win.

The problem: optimal on paper, mediocre in the real world

Mean-variance optimization is supposed to be strictly better than naive diversification. It uses information. It accounts for which assets are risky and which move together. Mathematically, an equal-weighted portfolio is almost never on the efficient frontier.

But here is what that argument quietly assumes: that you know the expected returns and covariances. You do not. You estimate them from historical data, and those estimates carry error. A large body of work, from Michaud to Best and Grauer to Chopra and Ziemba, had already shown that optimizers amplify these errors.

So the honest question is not "which portfolio is best if I knew the truth." It is: given that I must estimate my inputs from real, finite, noisy data, does optimization actually leave me better off than not bothering?

That is what this paper tested, and nobody had previously tested it this systematically.

The key idea via analogy: a race with a fair starting line

Think of it like comparing a Formula 1 car to a bicycle, but the race is not on a track. It is through a swamp. The F1 car is objectively the superior machine, but it needs a road, and there is no road. In the swamp, the bicycle keeps moving.

The authors set up the swamp carefully. They took fourteen different portfolio models, including plain sample-based mean-variance, minimum variance, Bayesian approaches, shrinkage estimators, and models with constraints, and they ran all of them across seven different datasets of real asset returns.

Critically, they ran everything out of sample. Each model gets a window of historical data, forms its portfolio, and then is scored on what happens next, on data it never saw. That is the only honest way to score a strategy that depends on estimated inputs. Fit a model in-sample and of course the fancy model wins, it has more knobs to fit the past with.

Then they compared everything to the 1/N benchmark, which uses no estimates at all. It does not need expected returns. It does not need a covariance matrix. It cannot suffer from estimation error, because it does not estimate anything.

The finding

Across the fourteen models and the seven datasets, none of the optimized strategies was consistently better than 1/N on the metrics that matter: Sharpe ratio, certainty-equivalent return, and turnover.

Sit with that. Decades of theory, and the naive rule holds its own. The gain from optimal diversification is real, but it is offset, and often more than offset, by the estimation error you take on to get it.

The authors then asked a wonderfully concrete follow-up: how much data would you need for the sophisticated approach to reliably win? Their answer is the number that people remember from this paper. For a portfolio of 25 assets, the estimation window required for sample-based mean-variance to beat 1/N is on the order of 3,000 months. For 50 assets, roughly 6,000 months.

Three thousand months is 250 years. And that is assuming the world stays statistically the same for those 250 years, which it obviously will not. In other words: for realistic numbers of assets and realistic data histories, you will never have enough data.

Why it mattered

• It changed the burden of proof. After this paper, if you propose a new portfolio optimization method, the first question you get asked is: does it beat 1/N out of sample? That is now the floor, not the ceiling. It is a devastatingly simple benchmark and a lot of methods fail it.
• It gave a rigorous foundation to a practical instinct. Practitioners had long defaulted to roughly equal-weighting, or heavily constrained portfolios, and were sometimes told they were being unsophisticated. This paper made the case that the instinct is statistically defensible.
• It clarified where the pain comes from. The estimation error in expected returns is the main killer. Strategies that avoid estimating expected returns entirely, like minimum-variance portfolios, performed relatively better. That observation helped fuel the whole low-volatility and risk-based investing movement.
• It is a general lesson about complexity. A model with more parameters is not more accurate, it is more parameter-hungry. If you cannot feed it, it starves in an expensive way. This is the same message as the bias-variance trade-off in machine learning, delivered in a portfolio context.

The honest limitations

• 1/N is not actually optimal, it is just hard to beat. The paper is not a proof that equal weighting is the right answer. It is a demonstration that the alternatives, as usually implemented, do not reliably improve on it. Those are different claims.
• The choice of assets does a lot of work. Equal-weighting across a set of broadly sensible, roughly comparable assets is a very different thing from equal-weighting across a set that includes one terrible asset. 1/N contains a hidden assumption that your menu is already reasonable. Someone chose that menu, and that choice is a portfolio decision the benchmark gets for free.
• The critics have a real point. Kritzman, Page and Turkington argued in 2010 that the failure here is driven by how the optimizers were fed, in particular the use of short rolling windows of historical returns as forecasts of future returns, which produces implausible inputs. With longer samples or more sensible assumptions, they found optimization does beat 1/N. So this is a live argument, not a settled verdict.
• Turnover and costs cut both ways. 1/N looks great partly because rebalancing to equal weights is cheap and stable. In a universe with thousands of names, or illiquid assets, equal weighting has its own practical problems.

The one-line takeaway

DeMiguel, Garlappi and Uppal showed that once you account for the estimation error you inevitably introduce by having to guess your inputs, the theoretically optimal portfolio usually fails to beat the theoretically naive one, and you would need centuries of data before it reliably would.