The Optimizer That Maximizes Your Mistakes: Michaud's Markowitz Enigma

Here is a puzzle that bothered Richard Michaud in the late 1980s, and it is still a good interview question today. Mean-variance optimization is mathematically correct. It provably gives you the best possible portfolio for a given set of inputs. It won Harry Markowitz a Nobel Prize. And yet, decades after it was invented, most professional money managers refused to use it.

Were the practitioners just too dumb to appreciate the math? Michaud's 1989 paper says no. The practitioners were right, and for a reason worth understanding.

The problem: the math is perfect, the inputs are not

The optimizer takes two things from you: your forecasts of expected returns, and your estimates of how assets move together. It then treats those numbers as if they were handed down from heaven. It does not know, and cannot know, that your expected return for a particular stock is really a wobbly guess made from noisy history.

That would be tolerable if the optimizer merely passed your uncertainty through unchanged. It does something much worse.

Ask yourself what makes an asset look attractive to an optimizer. Three things: a high expected return, low volatility, and low correlation with everything else in the portfolio. Now ask what an estimation error in the flattering direction looks like. It looks like an asset whose return you accidentally overestimated, or whose risk you accidentally underestimated, or whose diversification benefit you accidentally overstated.

Notice that those are the same thing. The assets that look most appealing to the optimizer are, disproportionately, the assets where your numbers happen to be wrong in the direction that flatters them. So the optimizer does not just tolerate your errors. It seeks them out and concentrates the portfolio in them.

Michaud gave this behavior the name that stuck: mean-variance optimizers are, in practice, "estimation-error maximizers." That single phrase is probably the most quoted line in the practical portfolio construction literature.

The key idea via analogy: hiring based on one lucky quarter

Imagine you run a sales team and you want to promote your best performer. You have one quarter of data. One rep sold far more than anyone else. The optimizer's logic is: promote that rep, and give them the whole territory.

But you know that one blowout quarter is a mix of genuine skill and luck. If you rank on one quarter and then hand everything to the top name, you are systematically promoting whoever got the luckiest, because among people of roughly similar skill, the one at the top of a short, noisy sample is usually the one with the best luck. Their performance will not repeat.

That is a mean-variance optimizer looking at asset returns. It ranks on a noisy sample, then piles into the top of the ranking, and the top of the ranking is enriched with luck.

This explains the specific symptoms practitioners had always complained about:

• Absurd concentration. The "optimal" portfolio often puts enormous weight in a couple of assets and zero in most others, which no sensible investor would accept.
• Wild instability. Nudge one expected return by a fraction of a percent and the whole portfolio reshuffles. A tool whose answers flip when the inputs barely move is not giving you a robust answer.
• Terrible out-of-sample results. The mathematically optimal portfolio frequently underperforms a naive one, because it was optimized for a world made of estimation noise that then failed to show up.

And there is a further sting. Michaud pointed out that the optimizer tends to concentrate in exactly the assets that a portfolio manager already knows the least about, because thinly-covered and unusual assets tend to have the noisiest estimates, and noisiest estimates produce the most extreme apparent opportunities.

Why it mattered

• It legitimized the practitioner's distrust. Before this paper, the standard view was that resistance to optimizers was a failure of professional education. Michaud reframed it as a well-founded response to a real statistical defect. That is a big deal in a field where academic tools are often adopted uncritically.
• It named the disease so people could treat it. Once you accept that estimation error is the enemy rather than an annoyance, an entire research agenda opens. Shrinkage estimators, Bayesian priors, portfolio constraints, robust optimization, resampling, and equal-weighting all became responses to the problem Michaud articulated. The Black-Litterman model, Ledoit-Wolf shrinkage, and the later "1/N is hard to beat" literature all live downstream of this diagnosis.
• It pushed the field toward uncertainty-aware portfolio construction. Michaud argued that the honest answer to "what is the optimal portfolio" is not a single point but a fuzzy region, because your inputs are estimates with error bars. He went on to develop a resampling-based approach in his later work, averaging optimal portfolios computed across many simulated versions of the inputs to get something more stable than any single run.

The honest limitations

• Diagnosis is easier than cure. The paper is far more persuasive about what breaks than about what to do instead. The remedies proposed in it and in the follow-up literature are all imperfect trade-offs.
• Resampling has its own critics. The resampled approach Michaud later championed has been challenged on theoretical grounds, including the objection that averaging portfolios from simulated inputs does not have a clean decision-theoretic justification, and that it can be sensitive to how you simulate. It is a heuristic that often works, not a proof.
• It is easy to over-learn the lesson. "Optimizers are estimation-error maximizers" got so popular that some people concluded optimization is useless and defaulted to equal weights. Kritzman, Page and Turkington later pushed back hard, arguing the problem is careless inputs (like short rolling windows of past returns as return forecasts), not optimization itself. Optimizers punish bad inputs, which is not the same as optimizers being bad.
• The paper is about a symptom of a deeper issue. The true problem is that expected returns are extremely hard to forecast. Any method that requires them, optimizer or not, inherits that difficulty.

The one-line takeaway

Michaud showed that a mean-variance optimizer is not a neutral machine that processes your forecasts, it is a machine that hunts for and concentrates in whichever of your forecasts happens to be most flatteringly wrong, which is why mathematically optimal portfolios so often turn out to be practically terrible.