What Is an Eigenportfolio?

An eigenportfolio is a portfolio whose weights are derived from an eigenvector of an asset return covariance or correlation matrix. Each eigenportfolio represents one orthogonal direction of variation in the selected return sample. The associated eigenvalue measures the variance captured by that direction.

The first eigenportfolio often resembles a broad common market movement because it captures the largest variance direction. Later eigenportfolios can resemble sector contrasts or long-short spreads. These interpretations come from the data and can change through time; an eigenvector has no economic label or expected-return direction by itself.

Mathematical definition

For an \(N\times N\) covariance matrix \(\Sigma\), eigenvalue \(\lambda_k\) and unit-length eigenvector \(\mathbf{v}_k\) satisfy

\[\Sigma\mathbf{v}_k = \lambda_k\mathbf{v}_k, \qquad \mathbf{v}_k^\top\mathbf{v}_k=1.\]

For a symmetric covariance matrix, the eigenvectors can be chosen orthonormal:

\[\mathbf{v}_j^\top\mathbf{v}_k=0 \quad\text{for }j\ne k.\]

If returns are centred and the eigenvector is used directly as a weight vector, the unit-norm eigenportfolio return is

\[f_{k,t}=\mathbf{v}_k^\top\mathbf{r}_t\]

with variance

\[\operatorname{Var}(f_k) = \mathbf{v}_k^\top\Sigma\mathbf{v}_k =\lambda_k.\]

An investable portfolio normally needs another normalization for gross, net, volatility, or capital exposure.

Two-asset worked example

Suppose two assets each have variance \(0.04\) and covariance \(0.03\):

\[\Sigma = \begin{bmatrix} 0.04 & 0.03\\ 0.03 & 0.04 \end{bmatrix}.\]

The eigenpairs are

\[\lambda_1=0.07, \qquad \mathbf{v}_1=\frac{1}{\sqrt{2}} \begin{bmatrix}1\\1\end{bmatrix},\]

and

\[\lambda_2=0.01, \qquad \mathbf{v}_2=\frac{1}{\sqrt{2}} \begin{bmatrix}1\\-1\end{bmatrix}.\]

The first direction moves both assets together and captures seven times the variance of the second. The second is a spread between them.

For a unit-gross presentation, divide each vector by the sum of its absolute weights. The portfolios become \((0.5,0.5)\) and \((0.5,-0.5)\). Their variances are \(0.035\) and \(0.005\), respectively; the rescaling halves each unit-norm variance because the weight vector is divided by \(\sqrt{2}\).

If the next returns are \(+4\%\) and \(+2\%\), the unit-gross common portfolio earns

\[0.5(4\%)+0.5(2\%)=3\%,\]

and the unit-gross spread earns

\[0.5(4\%)-0.5(2\%)=1\%.\]

Python calculation

import numpy as np

covariance = np.array(
    [
        [0.04, 0.03],
        [0.03, 0.04],
    ]
)

eigenvalue, eigenvector = np.linalg.eigh(covariance)
order = np.argsort(eigenvalue)[::-1]
eigenvalue = eigenvalue[order]
eigenvector = eigenvector[:, order]

# Eigenvector signs are arbitrary; orient the first asset positively for display.
eigenvector *= np.sign(eigenvector[0])
unit_gross_weight = eigenvector / np.abs(eigenvector).sum(axis=0)

print(np.round(eigenvalue, 2))
print(np.round(unit_gross_weight, 2))

Covariance PCA and correlation PCA

Covariance PCA gives more influence to assets with larger return variance. Correlation PCA first standardizes each return by its volatility and studies the dependence structure on a common scale. The two methods produce different eigenvectors.

When an eigenvector \(\mathbf{v}_k\) is estimated from a correlation matrix of standardized returns, a return portfolio in original asset units may use volatility-adjusted weights proportional to

\[w_{i,k}\propto\frac{v_{i,k}}{\sigma_i}.\]

The final normalization must then state gross, net, and volatility conventions. Calling both the correlation eigenvector and its volatility-adjusted asset weights “the eigenportfolio” without explanation creates avoidable ambiguity.

Positive and negative weights

Positive and negative eigenvector elements show which assets move on opposite sides of that statistical direction. The entire vector can be multiplied by \(-1\) without changing the eigenvalue or the represented subspace. Its sign is therefore arbitrary until an orientation convention is imposed.

Eigenvalues of a valid covariance matrix are non-negative. A materially negative estimated eigenvalue indicates that the matrix is not positive semidefinite, often because of inconsistent data, pairwise estimation, or numerical error.

Eigenportfolios as statistical factors

The leading eigenportfolios provide a compact statistical factor model. If \(K<N\) leading directions are retained, returns can be approximated by

\[\mathbf{r}_t \approx V_K\mathbf{f}_t+\boldsymbol{\varepsilon}_t,\]

where \(V_K\) contains the retained eigenvectors and \(\boldsymbol{\varepsilon}_t\) is residual return. The model separates common variation from asset-specific or lower-variance directions under the estimated sample covariance.

Residuals can be studied for relative-value or mean-reversion hypotheses, but PCA alone supplies no forecast. Any trading signal built from a residual needs its own timing, economic rationale, parameter estimation, and walk-forward evaluation.

Common implementation mistakes

Using future returns in the covariance window

Historical eigenportfolios must be estimated using returns available before the rebalance and then held fixed over the evaluation period.

Ignoring eigenvector sign and ordering instability

Signs can flip between runs, and nearby eigenvalues allow eigenvectors to rotate substantially. Align directions by subspace or a recorded orientation before measuring turnover.

Mixing covariance and correlation weights

Standardization changes the problem. Store whether PCA used covariance or correlation and how vectors were mapped back to asset weights.

Treating a principal component as alpha

An eigenportfolio describes realized covariance. Expected return requires a separate forecast and validation.

Retaining too many unstable components

Small eigenvalues and their vectors are especially sensitive to estimation noise. Shrinkage, sample-length checks, and out-of-sample stability matter.

Forgetting investability

Mathematical eigenvectors can be concentrated, illiquid, high-turnover, or expensive to short. Portfolio constraints alter their orthogonality and return representation.

Use in risk and portfolio evaluation

Eigenportfolios reveal dominant covariance directions, help diagnose hidden concentration, and can reduce the dimension of risk models. A portfolio's exposure to eigenportfolio \(k\) is the projection of its risk or holdings into that direction under the chosen convention.

Compare statistical factors with observable factor exposures such as market, sector, size, value, and crowding. A leading component may span several of them, while an economic factor may occupy several principal directions. The mapping can change as covariance changes.

Frequently asked questions

Is the first eigenportfolio the market portfolio?

It often resembles a broad common market direction in equity data, but this is an empirical interpretation. The result depends on the universe, return frequency, window, covariance or correlation choice, and normalization.

Are different eigenportfolios uncorrelated?

They are uncorrelated under the covariance matrix used to estimate them and the matching unit-norm construction. Realized out-of-sample correlations need not remain zero.

Why can eigenportfolio weights change suddenly?

Sampling noise, regime changes, universe changes, and close eigenvalues can rotate the estimated directions. Stable subspaces can exist even when individual vectors are unstable.

Can eigenportfolios be combined with fundamental factors?

Yes. Statistical directions can supplement observable exposures in risk analysis, provided overlap, stability, and interpretability are measured.

Next reading

• What Is a Factor-Mimicking Portfolio?
• What Is Residual Alpha?
• Factor exposure: security and portfolio sensitivity

Sources

• Marco Avellaneda and Jeong-Hyun Lee, “Statistical Arbitrage in the U.S. Equities Market”, Quantitative Finance, 2010.
• Karl Pearson, “On Lines and Planes of Closest Fit to Systems of Points in Space”, Philosophical Magazine, 1901.