Direct Answer
A covariance sensitivity tool shows how much a single covariance or correlation estimate changes the optimal portfolio weights produced by mean-variance optimization, even when the input error is small. Pick one covariance parameter, set a range for it, and the tool recomputes the minimum-variance portfolio weights across that range so the fragility becomes visible directly. This illustrates why optimizers are sensitive to small estimation errors, Michaud (1989) called this the "error maximizer" problem.
Covariance Sensitivity Tool
Estimation error in the covariance matrix causes large, unpredictable swings in optimal portfolio weights, even when the estimation error is small. This tool demonstrates that fragility directly: choose any single covariance (or correlation) parameter, set its range of variation, and see how the optimal minimum-variance portfolio weights respond across that range. Michaud (1989) called the optimizer an "error maximizer" for exactly this reason.
Tool Inputs
Weight Sensitivity Chart
Each line shows how one asset's optimal weight in the minimum-variance portfolio changes as the selected correlation varies. Large slopes = high sensitivity = fragility to estimation error in that parameter.
Sensitivity Table
Results will appear here.
Sensitivity Statistics
Run the analysis to see statistics.
What This Shows
- Steep slopes in the chart indicate high sensitivity: If the optimal weight for an asset changes by 20 percentage points when a correlation moves by 0.1, then a small estimation error in that correlation produces a large error in the portfolio weights.
- The range of motion over a realistic estimation error: A correlation estimated from T=300 observations (about 25 years of monthly returns) has a standard error of approximately 1/√T ≈ 0.058, while a 36 month window has a standard error of about 0.17. A sweep from 0.2 to 0.5 spans about five standard errors in total when T=300 (0.3 divided by 0.058), or roughly plus or minus two and a half standard errors around its midpoint, the kind of uncertainty that is always present in practice.
- Concentrated weight changes reveal the error-maximizer problem: When a single asset's weight swings between 0% and 60% over a plausible correlation range, the optimizer is not producing a robust portfolio. It is amplifying small estimation errors into large allocation changes.
- The fix is robust optimization: Shrinkage (Ledoit-Wolf), resampling, or adding weight constraints all reduce slope, they make the weight lines flatter. Constraints cap the maximum sensitivity by bounding how far weights can move. Shrinkage pulls the covariance matrix toward a target that has less extreme inverse eigenvalues, reducing the multiplier applied to estimation errors.
- Minimum-variance is typically more stable than max-Sharpe: Because min-variance doesn't use expected return estimates, it avoids the dominant source of estimation error. The slopes you see in this tool reflect covariance estimation error only; the max-Sharpe optimizer adds expected return estimation error on top, making it even more sensitive.
Frequently Asked Questions
What does the covariance sensitivity tool show?
It sweeps one correlation across the range you set, re-solves the optimal weights at each step and charts how far each asset's weight moves. A steep line means the allocation depends heavily on an input that is itself estimated with error; a flat line means the optimizer would give a similar answer whichever value the correlation really takes.
How large is the estimation error in a correlation?
For a rolling estimate built from T observations the standard error is roughly one over the square root of T, so a correlation estimated from about 300 data points carries a standard error near 0.06. A sweep from 0.2 to 0.5 therefore spans about five standard errors in total, or roughly plus or minus two and a half around its midpoint, which is the uncertainty that is always present in practice rather than an extreme case.
How do you make an optimized portfolio less sensitive to its inputs?
Weight constraints cap how far any weight can move, shrinkage pulls the covariance matrix toward a target with less extreme inverse eigenvalues, and resampling averages the solutions across many perturbed inputs; each flattens the weight lines in the chart. Minimum-variance portfolios are usually steadier than maximum-Sharpe ones because they do not use expected return estimates, the dominant source of instability.
References
- Markowitz, H. (1952). "Portfolio Selection." The Journal of Finance, 7(1), 77-91, the mean-variance framework whose sensitivity to covariance-matrix inputs this tool demonstrates.
- CFA Institute Research and Policy Center
- National Bureau of Economic Research: portfolio theory and asset pricing research
Educational Disclaimer
This tool is for educational purposes only. Results are illustrative of general properties of mean-variance optimization and covariance sensitivity, they do not represent any specific investment strategy or recommendation. The projected gradient optimizer used here may not find the global minimum for all input combinations. This tool does not constitute investment advice.