Michaud Resampled Efficiency

Michaud resampled efficiency addresses the “error maximization” behavior of mean-variance optimization. A conventional frontier uses one estimated mean vector and covariance matrix, then commits to the resulting sharp allocations. Resampling instead generates many plausible input sets, optimizes each, and averages the portfolios that occupy comparable frontier locations. The result is usually less concentrated and more stable.

The motivation is sound: inputs are estimates. The stronger claim—that resampling always creates a superior efficient frontier—does not follow automatically. Bootstrap design, frontier matching, constraints, and transaction costs determine the practical result.

The workflow

1. Estimate return-model parameters from a training window.
2. Draw simulated or bootstrap samples of returns.
3. Estimate μb, Σb in every resample.
4. solve a constrained Markowitz optimization

for each target risk or return.

1. Average weights across resamples at each target.
2. evaluate the averaged portfolios out of sample.

w̄_k = (1/B) Σ[b=1..B] w_(b,k)

Choice	Consequence
iid bootstrap	ignores serial dependence
block bootstrap	retains short-run dependence
parametric simulation	smooth but model-dependent
long-only bounds	often dominate resampling effect
target matching	can average incomparable portfolios

Why weights become smoother

The optimizer frequently switches to a different asset when estimates are perturbed. Weight averaging dilutes these unstable corner solutions. It behaves like a nonlinear regularizer, although it does not explicitly state a preference for diversification. That makes it less transparent than shrinkage or a direct L2 penalty.

import numpy as np

def stationary_bootstrap_indices(n, block_mean, rng):
    idx, out = rng.integers(n), []
    for _ in range(n):
        out.append(idx)
        if rng.random() < 1 / block_mean:
            idx = rng.integers(n)
        else:
            idx = (idx + 1) % n
    return np.array(out)

# For each draw: estimate inputs, solve the same constrained optimizer,
# then average weights only after aligning target volatility or return.

The illustration deliberately omits the optimizer. Its constraints must be identical across resamples; otherwise the average reflects changing mandates rather than parameter uncertainty.

Important critiques

Resampling historical observations does not fix a biased alpha model, a regime break, or an implausible covariance. It can also obscure the actual decision rule: the averaged portfolio may not be mean-variance efficient under the original estimates. In addition, naive averaging may have high turnover when re-estimated monthly, especially if the frontier target is chosen by in-sample Sharpe.

Compare resampling honestly with simpler controls:

Alternative	What it makes explicit
covariance shrinkage	distrust of sample correlations
Bayesian/Black-Litterman	prior and view confidence
robust optimization	uncertainty set
risk parity	no alpha forecast dependence
weight penalties	diversification and turnover preference

See covariance shrinkage for a low-complexity baseline that often captures much of the stability benefit.

Validation design

Use nested walk-forward testing: choose the number of resamples, bootstrap method, and frontier point using only prior data. Rebalance on a realistic schedule and deduct costs. Record realized volatility, concentration, drawdown, turnover, and forecast error—not only return. Resampling is useful if it improves these jointly across market regimes, not if it merely creates a prettier in-sample frontier.

For multi-asset portfolios, preserve factor and duration exposures in the comparison. A low-turnover resampled allocation that silently changes its equity beta is not a clean improvement over the base portfolio.

Report dispersion across resampled allocations as well as their mean. Wide percentile ranges identify holdings whose apparent averaged precision comes from cancellation of incompatible optimizer outcomes rather than a stable investment conclusion.

Keep the baseline estimator, constraints, and reporting schedule fixed. Otherwise, resampling may look successful simply because it smuggles in a more conservative implementation than the portfolio it is supposed to evaluate.

Key takeaways

• Resampled efficiency averages frontier allocations across plausible estimated inputs.
• It often reduces corner weights because unstable optimizer choices cancel out.
• Bootstrap assumptions and target alignment are part of the model, not implementation details.
• Compare it with explicit shrinkage, robust constraints, and cost-aware baselines out of sample.