Start with the same observations
Compare returns for the same dates, in the same currency and with the same payout accounting. Price and yield are different series; the correlation of rising price levels can describe a general trend rather than the desired short-term dependence. Omissions and different trading hours can shift the apparent relationship even between economically related assets.
Determine the frequency: Daily, monthly, and annual changes produce different sets of observations. A long window adds history, but can mix different modes. The short one notices changes faster, but is more sensitive to randomness. There is no universal length that makes future correlation known. Save the method and check how the output changes with reasonable window changes.
Model of two training parts
Let the weights of the conditional portfolio be sixty and forty percent. Volatility for the same period is fifteen and twenty-five percent, correlation is 0,2. In decimal notation, the covariance is 0,2 × 0,15 × 0,25, that is, 0,0075. This is not a percentage return, but a joint dispersion value at the appropriate scale.
The portfolio variance is obtained as 0,6² × 0,15² + 0,4² × 0,25² + 2 × 0,6 × 0,4 × 0,0075 = 0,0217. The square root gives a volatility of about 14,73%. If, with the same weights and individual volatilities, the correlation is one, the result is nineteen percent. All parameters are specified and not estimated based on the real market; shares are not a recommendation.
Why does a crisis change the meaning of a quiet window?
The calm period estimate may not reflect the joint response to liquidity shortages, a common source of funding, or changing expectations. Low historical connection does not give a contractual obligation to move independently. Concentration may be hidden by the similar dependence of different securities on one group or industry; FINRA recommends looking at such connections as well.[103](#source-cv-concentration)
Be especially careful when comparing assets with rare revaluations. The stale price gives the impression of low volatility and low correlation, although the actual sale is possible at a deep discount. The absence of daily transactions is not a separate source of protection. Statistical smoothness may indicate the measurement method rather than the properties of the asset.
What's left out of the formula?
The variance formula does not automatically account for default, lockout, non-linear payouts, and permanent loss of value. It does not show the maximum drawdown and does not determine the probability of loss without additional distributional assumptions. A reduction in one statistical indicator does not mean a reduction in all possible risks.
The working check contains the original series, cleaning rules, currency, payout accounting, window and individual stress combinations. If the conclusion about the distribution of risk disappears with a small change in the period, this should be shown rather than choosing the most favorable coefficient. The result is an assessment of historical linear dependence with clear boundaries. It helps ask questions about concentration, but does not guarantee future independence or set the right portfolio for the person.
Calculation conditions are more important than extra signs
Indicate whether sample or other covariance estimation is used and how incomplete observations are handled. When comparing multiple assets, pairwise missingness removal may mean that each relationship is measured on a different set of dates. Then the final matrix requires additional compatibility checking; it cannot be considered automatically consistent just because each cell lies between the acceptable correlation limits.
For a training example, it is enough to keep the same observations and a transparent formula. For a large set, it is useful to check how the conclusions change when individual periods are excluded and expenses are added. Such an analysis does not predict the next crisis, but shows the dependence of the result on the method of measurement.
Finally, correlation does not distinguish between causation. Two rows may respond to a common external factor rather than to each other. Therefore, an economic description of connections remains necessary even after statistical calculations. It helps to build stress that has not yet been present in history, and not to mistake a mathematically beautiful matrix for a complete map of risks.
**Related Basics:** [Diversification: look for independent causes rather than counting names](/guides/diversification/); [Portfolio risk: losses, timing and joint movement of assets](/guides/portfolio-risk/).
Expectation and the efficient frontier - the next question after variance
The average expected return of the portfolio is equal to the sum of the weights multiplied by the expected returns of the assets: μp = Σ wiμi. This is the mathematical expectation for the agreed period, and not the promised result. In Sharpe's account, the efficient set excludes combinations for which there is an alternative with no less expected return and no greater risk for at least one strict improvement.[1](#source-cp-efficient-frontier) Here, the risk within the model means the standard deviation, not the entire set of consequences for the person.
Let's define two fictitious assets A and B, without debt and short positions. Their expected annual returns are 6% and 10%, both volatilities are 20%, and there is zero correlation. Let us denote the weight of B as w, the weight of A as 1 − w. Then μp = 6% + 4% × w, and σp² = 0,04 × [(1 − w)² + w²]. The parameters are set for the lesson, not obtained from a historical sample.
| Weight B | Expected return | Standard Deviation | |---|---|---| | 0% | 6% | 20,00% | | 25% | 7% | 15,81% | | 50% | 8% | 14,14% | | 75% | 9% | 15,81% | | 100% | 10% | 20,00% |
The mixture with weight B 25% is inferior to the mixture with weight B 75%: the standard deviation is the same, the expectation is lower. The minimum variance is achieved at 50%. In this continuous two-active set, the effective branch goes from weight B 50% to 100%; smaller weights are dominated by symmetrical larger ones. To get the expected 9%, you need w = (9 − 6)/4 = 0,75, but this is only fulfilling a mathematical constraint, not choosing the appropriate yield. On the graph, σp is plotted horizontally and μp vertically; the effective bound is the upper left part of the feasible set, not the entire bend.
A small expectation error changes the optimal weight
A single example reveals sensitivity. Let the volatility again be 20%, but the correlation is 0,95: the assets are very similar. We maximize the conditional criterion J = μp − (λ/2)σp² with λ = 4 and weights from zero to one. This is a selected instructional trade-off function, not a measured reader utility. With equal expectations 8%, the optimum is 50% B.
For this symmetric problem, the internal solution w = 0,5 + (μB − μA) / [2λσ²(1 − ρ)], then it is limited by the segment [0; 1]. The denominator is 0,016. If B's expectation is raised from 8% to 8,4%, leaving everything else unchanged, the model will produce 75% B. If 8,8% is 100%. Changing the score to 0,4 percentage points changed the weight to 25 percentage points: this is not the same change.
The reason is not knowing the future winner. With similar assets, the penalty for weight shifting is small, so the noisy difference in expectations becomes decisive. The historically calculated expectation does not become accurate until the optimizer is run. For a different covariate or criterion, the results will be different; minimum variance itself does not use μ at all.
Compare weights across multiple reasonable estimates, introduce explicitly stated concentration and turnover limits, and review costs and individual access scenarios. These restrictions may reduce sensitivity, but do not prove the correctness of the input data. Efficiency refers to a given set, period and criterion. The model does not imply either the safety or stability of the optimal shares, or the advantage of the settlement strategy after expenses.