Resampled Efficiency

A plain mean-variance optimizer takes your best guesses for returns and risks and hands back a single "optimal" portfolio. The trouble, covered in Pitfalls of Mean-Variance Optimization, is that those guesses are noisy, and the optimizer treats every decimal place as gospel. Change one asset's estimated return by a fraction of a percent and the "optimal" weights can swing from a big long to a big short. Resampled efficiency, invented by Richard Michaud, is a way to stop the optimizer from taking your noisy estimates so literally.

The idea is disarmingly simple: instead of trusting one set of estimates, admit you don't know them exactly, generate many alternative sets that are all consistent with your data, optimize a portfolio for each, and then average the portfolios. The lucky over-bets in one scenario get cancelled by the unlucky ones in another, and what survives the averaging is the part of the allocation that holds up across many plausible worlds.

How it works, step by step

You start with your point estimates of the mean-return vector $\mu$ and covariance matrix $\Sigma$. Then you repeat the following many times (say 500):

1. Simulate. Draw a fresh, fake return history of length $T$ from your estimated $\mu$ and $\Sigma$. This is a Monte Carlo bootstrap (Bootstrap and Resampling): each draw is a world that could plausibly have produced your data.
2. Re-estimate. Compute new sample estimates $\hat\mu, \hat\Sigma$ from that fake history.
3. Optimize. Trace the The Efficient Frontier for those estimates and record the optimal weights at each rung of risk (or each return target).

After all the runs, you average the weight vectors at each rung across every simulation. That averaged set of portfolios is the resampled frontier.

$$w^{\text{RE}}(k) = \frac{1}{B}\sum_{b=1}^{B} w^{(b)}(k),$$

where $B$ is the number of simulations, $k$ indexes the risk rung, and $w^{(b)}(k)$ is the optimal weight vector from simulation $b$. In words: the resampled portfolio at each risk level is just the average of all the optimizers' answers for that level.

Worked example

Take two nearly identical assets with equal true expected returns, equal 20% volatility, and correlation 0.9. Suppose one sample happens to give estimated returns of 8.3% and 7.7%. A single mean-variance run reads that 0.6% gap as a near-arbitrage between twins and prescribes something wild, roughly 250% long the "winner" and 150% short the "loser."

Now resample. Across 500 simulated histories, the winner and loser trade places about half the time, because the true means are equal. When you average all 500 optimal weight vectors, the giant long and short cancel out, and the resampled allocation lands near a sensible 50/50 split. Same inputs, but the answer stopped chasing noise. Out of sample, that 50/50 book keeps far more of its Sharpe and turns over a fraction as much.

Where it helps and where it misleads

Resampling shines because it regularizes without you having to pick a penalty: no-short and position limits emerge naturally from averaging, and turnover collapses because the weights barely move as new data arrives. It is closely related to shrinkage (Ledoit-Wolf Covariance Shrinkage) and to Bayesian averaging (The Black-Litterman Model), all three pull the aggressive point solution back toward something humbler.

The honest caveats: resampled efficiency has no clean closed-form justification the way The Black-Litterman Model does, it can leave tiny leftover weights in assets that should be zero, and the method is patented, which limited its academic adoption. Still, as a practical antidote to the error-maximizing optimizer, it works, and the intuition, don't trust one draw when you can average many, is worth carrying into any estimation problem.