The short answer
Markowitz (1952) rejects the maximize-expected-return rule because it never implies diversification, and replaces it with the expected-return–variance-of-returns rule. On a worked stocks/bonds/gold tape, the minimum-variance portfolio's 3.56% volatility falls below the 4.25% of its calmest asset, the covariance term made literal.
Markowitz's 1952 Portfolio Selection did one thing no prior framework did: it made diversification a mathematical consequence of variance and covariance, not a folk maxim. The paper rejects the rule "maximize expected return" because that rule never tells you to hold more than one security, and it replaces it with the expected-return–variance-of-returns (E-V) rule. Worked here on a real three-asset tape through the Efficient Frontier Builder: a stocks/bonds/gold blend whose minimum-variance portfolio carries 3.56% annualized volatility, below the 4.25% of its lowest-vol constituent. That gap is the covariance term doing its job, and it is the entire point of the 1952 paper.
What Markowitz actually rejected
The opening move of the paper is a rejection. Markowitz (1952) argues that the rule by which "the investor does (or should) maximize discounted expected, or anticipated, returns" must be discarded, because it "fails to imply diversification no matter how the anticipated returns are formed." If the only goal is maximum expected return, the rule tells you to put 100% into whichever single security has the highest expected return. That is the opposite of what every sensible investor does.
The fix is the expected-return–variance-of-returns rule (the E-V rule): the investor "should diversify and ... maximize expected return," treating expected return as a thing to want and variance of return as a thing to avoid. Risk enters the objective as a first-class quantity. This is the conceptual hinge of modern portfolio theory.
Why variance, not expected return, is where diversification lives
For a portfolio with weights w_i on assets with returns R_i, the expected return is a simple weighted average:
E[R_p] = Σ w_i · E[R_i]
Diversification does nothing here. The portfolio's expected return is always between the worst and best asset's expected return, no matter how you weight.
Variance is different. The portfolio variance is:
Var[R_p] = Σ_i Σ_j w_i · w_j · Cov(R_i, R_j)
= Σ_i w_i² · σ_i² + Σ_{i≠j} w_i · w_j · σ_i · σ_j · ρ_ij
The cross terms carry the correlations ρ_ij. When two assets are imperfectly correlated (ρ_ij < 1), the cross terms are smaller than they would be under perfect correlation, and the portfolio variance falls below the weighted average of the individual variances. Markowitz states the consequence directly: diversification across securities that are not perfectly correlated reduces variance, and it "is both observed and sensible." A rule that ignores covariance, he notes, "implies that the investor should diversify all his funds among securities which give maximum expected return" only when those securities are perfectly correlated, which they are not.
The worked tape
Three assets, 96 daily observations each, risk-free rate 3% annual: US Stocks, Treasury Bonds, Gold. Feeding the tape to the Efficient Frontier Builder returns the annualized moments and the two canonical portfolios. (Exact inputs and the full result are in the verified-output block below; the numbers in the prose are read directly from that run.)
| Asset | Annualized return | Annualized volatility |
|---|---|---|
| US Stocks | 11.00% | 15.14% |
| Treasury Bonds | 3.50% | 4.25% |
| Gold | 6.50% | 13.53% |
The pairwise correlations from the same run:
| Pair | Correlation |
|---|---|
| Stocks–Bonds | −0.311 |
| Stocks–Gold | +0.150 |
| Bonds–Gold | +0.019 |
The negative stocks–bonds correlation is the workhorse. It is the empirical pattern Markowitz's covariance term is built to exploit: when two assets tend to move in opposite directions, holding both shrinks the joint variance more than the weighted-average variance would suggest.
The minimum-variance portfolio: the free lunch made literal
The engine's minimum-variance portfolio weights are 12.24% Stocks, 83.37% Bonds, 4.39% Gold, with annualized volatility 3.56%.
Read that volatility again. It is lower than the 4.25% standalone volatility of Treasury Bonds, the least volatile asset in the universe. Holding a basket that is 83% bonds plus a slug of much-more-volatile stocks and gold produces a portfolio less volatile than pure bonds. The weighted average of the three standalone volatilities at these weights is 5.99%; the actual portfolio volatility is 3.56%. The 2.43-percentage-point gap is the covariance term, and the negative stocks–bonds correlation is most of it.
This is the result that makes the 1952 paper foundational. "Don't put all your eggs in one basket" is folk wisdom; "the minimum-variance basket is less volatile than its calmest ingredient" is a theorem, and it falls straight out of the quadratic form for portfolio variance.
The tangency portfolio: the best risk-adjusted mix
The minimum-variance portfolio is the leftmost point on the frontier, not the best one to hold. The tangency portfolio maximizes the Sharpe ratio over the risk-free rate. The engine returns weights 32.29% Stocks, 57.80% Bonds, 9.91% Gold, with annualized return 6.22%, volatility 5.13%, and Sharpe 0.627.
Compare that to the alternatives the same run lets you compute:
| Portfolio | Ann. return | Ann. vol | Sharpe |
|---|---|---|---|
| US Stocks only | 11.00% | 15.14% | 0.528 |
| Equal-weight (⅓ each) | 7.00% | 7.10% | 0.563 |
| Minimum-variance | 4.55% | 3.56% | 0.435 |
| Tangency (max-Sharpe) | 6.22% | 5.13% | 0.627 |
The highest-expected-return single asset (Stocks, 11%) has a worse Sharpe than the tangency mix. The naive "maximize expected return" rule would hold 100% stocks at Sharpe 0.528; the E-V rule finds a three-asset blend at Sharpe 0.627, more return per unit of risk, with a third of the volatility. That delta is precisely what Markowitz argued the older rule throws away.
The efficient set
Between the minimum-variance point and the high-return end, the engine sweeps 61 frontier points (here, return-volatility pairs running from roughly 1.2% return at 8.21% volatility on the lower branch up to 7.89% at 8.21% on the upper branch, with the minimum-variance vertex at 4.55% / 3.56%). Markowitz called the upper branch, the set of portfolios with maximum expected return for each level of variance, the efficient set. Every portfolio below the minimum-variance return at a given volatility is dominated: another portfolio offers the same volatility with a higher expected return.
The investor's job, in the 1952 framing, is two-stage. The first stage forms beliefs about the future (the expected returns, variances, and covariances). The second stage applies the E-V rule mechanically to pick a point on the efficient set, where on the frontier depends on the investor's risk appetite, but it is always on the frontier, never below it. The paper is explicit that it addresses only the second stage; forming the moment estimates is the harder, separate problem.
The catch the paper already names
The efficient set is only as good as the moments fed into it. Markowitz was clear that the E-V rule operates on beliefs about expected returns and covariances, not on known truths. On a 96-observation tape the estimation error in the annualized means is large, often the same order of magnitude as the means themselves, which is why mean-variance optimization on short, noisy data can produce extreme or unstable weights. The teaching tape here is engineered to be well-behaved (recentered to clean target means with a stable covariance structure) so the mechanics are visible. Real tapes are messier, and that messiness is the entire subject of the seventy years of literature that followed: shrinkage estimators (Ledoit-Wolf), Bayesian priors (Black-Litterman), resampling, and the robustness checks that keep a tangency solver from going haywire on thin data.
None of that revises the 1952 result. It refines the inputs. The E-V rule, the covariance-driven variance reduction, and the efficient set are exactly as Markowitz derived them.
What to take from the paper
- Maximizing expected return is not an investment rule. It never recommends diversification, which is the one thing real investors universally do.
- Variance is the home of diversification. Covariance terms let a portfolio be less volatile than its least-volatile member, demonstrated above where a 3.56% portfolio undercuts a 4.25% bond.
- Hold a portfolio on the efficient set. Anything else is dominated: same risk, less return.
- The frontier is only as honest as the moments. Garbage means in, garbage weights out. The 1952 math is sound; the estimation problem it sits on is where the work is.
Connects to
- Risk Parity vs Efficient Frontier: 3-Asset Portfolio Build: what happens when the same engine meets a noisy tape and the tangency solver goes short an asset.
- Risk-Adjusted Returns: Sharpe, Sortino, and Calmar on the portfolios the frontier produces.
- Correlation Matrix Visualizer: inspect the covariance structure the frontier consumes before you trust the weights.
- Conviction-Scaled Kelly Bet Sizing: once you have the mix, how much of the bankroll to commit.
- Efficient Frontier Builder: run your own tape.
- Efficient Frontier Builder methodology: the full input/output specification.
References
- Markowitz, H. M. (1952). "Portfolio Selection." The Journal of Finance 7(1), 77–91. DOI: 10.1111/j.1540-6261.1952.tb01525.x. JSTOR 2975974
- Markowitz, H. M. (1959). Portfolio Selection: Efficient Diversification of Investments. Wiley. The book-length treatment of the 1952 paper.
- Rubinstein, M. (2002). "Markowitz's 'Portfolio Selection': A Fifty-Year Retrospective." The Journal of Finance 57(3), 1041–1045. DOI: 10.1111/1540-6261.00453.
- Ledoit, O., & Wolf, M. (2004). "Honey, I Shrunk the Sample Covariance Matrix." The Journal of Portfolio Management 30(4), 110–119. DOI: 10.3905/jpm.2004.110.
- Black, F., & Litterman, R. (1992). "Global Portfolio Optimization." Financial Analysts Journal 48(5), 28–43. DOI: 10.2469/faj.v48.n5.28.
Verified engine output
Show the recompute-verified inputs and outputs
| name | |
|---|---|
| returns › row 1 | 0.01 |
| returns › row 2 | 0.02 |
| returns › row 3 | -0.005 |
| risk_free_annual | 0.03 |
| steps | 60 |
| assets › row 1 › name | US Stocks |
| assets › row 1 › returns (96 items) | [...] |
| assets › row 2 › name | Treasury Bonds |
| assets › row 2 › returns (96 items) | [...] |
| assets › row 3 › name | Gold |
| assets › row 3 › returns (96 items) | [...] |
| columns › row 1 | US Stocks |
|---|---|
| columns › row 2 | Treasury Bonds |
| columns › row 3 | Gold |
| n | 96 |
| m | 3 |
| mu ann › row 1 | 0.10999583650188005 |
| mu ann › row 2 | 0.03499945674539284 |
| mu ann › row 3 | 0.06501286364073988 |
| cov ann › row 1 › row 1 | 0.022923522304172377 |
| cov ann › row 1 › row 2 | -0.0020055347845657876 |
| cov ann › row 1 › row 3 | 0.0030658017424815783 |
| cov ann › row 2 › row 1 | -0.0020055347845657876 |
| cov ann › row 2 › row 2 | 0.0018100864018263155 |
| cov ann › row 2 › row 3 | 0.00010659859438421048 |
| cov ann › row 3 › row 1 | 0.0030658017424815783 |
| cov ann › row 3 › row 2 | 0.00010659859438421048 |
| cov ann › row 3 › row 3 | 0.01830597893117368 |
| min var › weights › row 1 | 0.12238801068327923 |
| min var › weights › row 2 | 0.8336821130758095 |
| min var › weights › row 3 | 0.04392987624091123 |
| min var › ret ann | 0.04549659972271777 |
| min var › vol ann | 0.03561272394445529 |
| min var › sharpe ann | 0.43514221902507716 |
| tangency › weights › row 1 | 0.3229168845006613 |
| tangency › weights › row 2 | 0.5779955329757599 |
| tangency › weights › row 3 | 0.0990875825235787 |
| tangency › ret ann | 0.0621910099777426 |
| tangency › vol ann | 0.051327993457037174 |
| tangency › sharpe ann | 0.6271628366826241 |
| frontier (61 items) | [...] |
| rf annual | 0.03 |
Computed live at build time.
Frequently asked questions
- What does Markowitz's 1952 paper actually prove?
- That diversification is a mathematical consequence of variance and covariance, not just folk wisdom. The paper rejects the rule of maximizing expected return because it never recommends holding more than one security, and replaces it with the expected-return–variance-of-returns (E-V) rule, where the investor maximizes expected return while minimizing variance. Portfolio variance carries covariance cross-terms, so an imperfectly-correlated basket can be less volatile than the weighted average of its parts.
- Why can a portfolio be less volatile than its least-volatile asset?
- Because portfolio variance includes covariance terms, not just individual variances. When two assets are negatively correlated, their cross-term is negative and subtracts from the joint variance. In the worked example, a basket that is 83% bonds plus stocks and gold has 3.56% annualized volatility, below the 4.25% standalone volatility of bonds, the calmest asset. The negative stocks-bonds correlation of -0.311 is most of the reason.
- Is the efficient frontier reliable on a short data tape?
- The frontier math is exact, but it operates on estimated means and covariances, and on a short tape the estimated means carry large error, often the same order of magnitude as the means themselves. That is why mean-variance optimization on thin or noisy data can produce extreme or unstable weights. Markowitz framed forming the moment estimates as a separate, harder problem; later work (shrinkage, Bayesian priors) refines the inputs without revising the 1952 result.
- What is the difference between the minimum-variance and tangency portfolios?
- The minimum-variance portfolio is the single lowest-volatility point on the frontier, ignoring expected return. The tangency portfolio maximizes the Sharpe ratio over the risk-free rate, the best return per unit of risk. In the worked example the tangency mix (32% stocks, 58% bonds, 10% gold) posts Sharpe 0.627, beating both the minimum-variance portfolio (0.435) and holding 100% of the highest-return asset (0.528).