Estimating a covariance matrix from historical prices is a core task in quantitative finance. The go-to solution is an exponentially weighted moving (EWM) estimate of the covariance matrix. It’s a simple, fast, and widely available method.
I use an EWM when I need something simple to get started, but there are a few things we can do to improve the estimates. In this article, we’ll explore three key ideas that can significantly improve covariance estimation:
-
Measurement - Instead of relying solely on backtest performance to evaluate covariance estimates, we can define objective metrics to assess their statistical fitness in isolation.
-
Decoupling variance and correlation - Variance and correlation behave differently over time and should be modeled independently.
-
Shrinkage - Pulling our noisy empirical estimates toward a structured target offers a systematic way to reduce estimation error.
This article won’t cover the various methods of covariance matrix estimation, such as EWM, GARCH or other methods. We’ll focus on the above three ideas to improve the estimates we get from these methods and we’ll use EWM estimates as the example.
Measurement
The most straightforward way to evaluate covariance matrix estimates is by backtesting. Simply run your complete investment strategy with different covariance estimators and compare the results.
However, this approach has some drawbacks. Backtesting can be time-consuming, making it impractical for rapid iteration and research. More importantly, mixing covariance estimation research with portfolio construction can lead to overfitting. When performance improves, you don’t know whether the gains come from better covariance estimates or inadvertent optimization of other portfolio components.
To address these issues, we need isolated evaluation methods that test covariance estimators independently of the full investment process.
The fiddly thing is that there is no “best” covariance matrix. We have to pick a general purpose metric, or, something that targets our particular need. In his book, Gappy goes into depth on different ways of measuring covariance matrices1.
Here, we’re going to cover two approaches; minimum-variance portfolios and the log-likelihood.
Minimum-variance portfolios
The idea here is that we usually want a covariance matrix to minimise the variance of our portfolio. We can construct the minimum-variance portfolio (MVP) using different covariance estimators and compare their realized variance. The estimator producing the portfolio with the lowest variance is considered the better option.
The MVP is given by: $$ \begin{aligned} \underset{\boldsymbol{w}_{t-1}}{\text{min}} &\quad \boldsymbol{w}^T_{t-1}\hat{\boldsymbol{\Omega}}_{t-1}\boldsymbol{w}_{t-1} \\ \text{s.t.} &\quad \boldsymbol{1}^T\boldsymbol{w}_{t-1} = 1 \end{aligned} $$ Where $\boldsymbol{w}_{t-1}$ are the portfolio weights calculated at time $t-1$, $\hat{\boldsymbol{\Omega}}_{t-1}$ is the covariance matrix of asset returns at time $t-1$ and $\boldsymbol{1}$ is a vector of ones. The constraint ensures that the weights sum to one.
The solution is: $$ \boldsymbol{w}_{t-1} = \frac{\hat{\boldsymbol{\Omega}}_{t-1}^{-1}\boldsymbol{1}}{\boldsymbol{1}^T\hat{\boldsymbol{\Omega}}_{t-1}^{-1}\boldsymbol{1}} $$ We can gain some intuition by imagining that the assets are uncorrelated. This means the covariance matrix $\hat{\boldsymbol{\Omega}}$ is a diagonal matrix of the variances. Then you can see that the portfolio weights are proportional to the inverse of the variances. This means that assets with lower variance will receive larger weights in the portfolio. Thereby dragging the portfolio variance down.
Here is some Python code to calculate it:
import numpy as np
import pandas as pd
def mvp(cov: pd.DataFrame) -> pd.Series:
"""
Returns the minimum-variance portfolio weights
from a covariance matrix.
Parameters
----------
cov : DataFrame
Covariance matrix of asset returns over
time.
"""
try:
icov = np.linalg.pinv(cov.values)
ones = np.ones(cov.shape[0])
w = icov @ ones / (ones @ icov @ ones)
except np.linalg.LinAlgError:
w = np.full(cov.shape[0], np.nan)
return pd.Series(w, index=cov.columns)
Once we have the portfolio weights, we can calculate the portfolio returns. Let’s say that the vector of asset returns at time $t$ is $\boldsymbol{r}_t$. The portfolio’s return at time $t$ is then: $$ p_t = \boldsymbol{w}^T_{t-1}\boldsymbol{r}_t $$ And the variance of the portfolio returns is: $$ \text{var}(p_1, p_2, \dots, p_T) $$ When comparing different estimators, we can calculate the variance of the portfolio returns for each estimator and select the one that produces the lowest variance.
The code for this metric is:
def mvp_metric(
covs: pd.DataFrame,
returns: pd.DataFrame,
) -> float:
"""
Returns the variance of the minimum-variance
portfolio.
Parameters
----------
cov : DataFrame
Covariance matrices of asset returns over
time. The matrix at time t should be the
matrix we can use to act on returns at
time t. Expected to have a column or index
called 'Date' that contains the date of
the matrix.
returns : DataFrame
Asset returns over time.
"""
portfolios = covs.groupby('Date').apply(mvp)
preturns = (portfolios * returns).sum(1, skipna=False)
var = np.var(preturns)
return var
Log-likelihood
Assume returns follow a multivariate Gaussian distribution. Under that assumption, we can use the likelihood function to evaluate how well each estimated covariance matrix explains the observed returns.
The likelihood that a vector of returns $\boldsymbol{r}_t$ is observed from a distribution with mean $\boldsymbol{\mu}_{t-1}$ and covariance $\hat{\boldsymbol{\Omega}}_{t-1}$ is given by the probability density function: $$ \text{P}(\boldsymbol{r}_t | \boldsymbol{\mu}_{t-1}, \hat{\boldsymbol{\Omega}}_{t-1}) = \frac{1}{(2\pi)^{n/2}|\hat{\boldsymbol{\Omega}}_{t-1}|^{1/2}} \exp\left(-\frac{1}{2}(\boldsymbol{r}_t - \boldsymbol{\mu}_{t-1})^T\hat{\boldsymbol{\Omega}}_{t-1}^{-1}(\boldsymbol{r}_t - \boldsymbol{\mu}_{t-1})\right) $$ Where $n$ is the number of assets.
We could iterate over all $t$ and multiply the likelihoods together to estimate the fit of our estimates: $$ \prod_{t} \ \text{P}(\boldsymbol{r}_t | \boldsymbol{\mu}_{t-1}, \hat{\boldsymbol{\Omega}}_{t-1}) $$ However, this can lead to numerical problems in practice. Instead, we convert the product into a sum by taking the logarithm of $\text{P}(\boldsymbol{r}_t | \boldsymbol{\mu}_{t-1}, \hat{\boldsymbol{\Omega}}_{t-1})$. This is known as the log-likelihood: $$ \log L(\boldsymbol{r}_t) = -\frac{1}{2} \left[ \log(|\hat{\boldsymbol{\Omega}}_{t-1}|) + (\boldsymbol{r}_t - \boldsymbol{\mu}_{t-1})^T\hat{\boldsymbol{\Omega}}_{t-1}^{-1}(\boldsymbol{r}_t - \boldsymbol{\mu}_{t-1}) + n \log(2\pi) \right] $$ We want the sum of these for all $t$ to be as large as possible.
We can simplify the log-likelihood function:
- For our purposes, we can assume that $\boldsymbol{\mu}_{t-1} = 0$.
- As we are optimising over the same set of $n$ assets, the term $n \log(2\pi)$ has no impact on our results. We can drop it.
- Multiplying by $-\frac{1}{2}$ only has the effect of flipping the sign, we can drop this too.
The final metric is: $$ \sum_t \left( \log(|\hat{\boldsymbol{\Omega}}_{t-1}|) + \boldsymbol{r}_t^T\hat{\boldsymbol{\Omega}}_{t-1}^{-1}\boldsymbol{r}_t \right) $$ When comparing different covariance estimators, we can calculate this metric for each estimator and select the one that produces the smallest value.
The code for this metric is:
def ll_metric(
covs: pd.DataFrame,
returns: pd.DataFrame
) -> float:
"""
Returns a log-likelihood based metric of the
fitness of the covariance matrices.
Parameters
----------
cov : DataFrame
Covariance matrices of asset returns over
time. The matrix at time t should be the
matrix we can use to act on returns at
time t. Expected to have a column or index
called 'Date' that contains the date of
the matrix.
returns : DataFrame
Asset returns over time.
"""
ll = np.full(len(returns), np.nan)
for i, date in enumerate(returns.index):
r = returns.loc[date]
cov = covs.loc[date].loc[r.index, r.index]
if cov.isnull().any().any():
continue
det = np.linalg.det(cov)
icov = np.linalg.inv(cov)
ll[i] = np.log(det) + r @ icov @ r.T
return np.nanmean(ll)
Example
Let’s walk through an example of how to use these metrics to compare different covariance estimators. We’ll vary the half-life parameter of an EWM covariance matrix and evaluate the results across a small set of ETFs.
Each ETF represents a distinct asset class:
-
SPY – U.S. equities (S&P 500)
-
TLT – Long-term U.S. Treasury bonds
-
GLD – Gold
-
GSG – Broad commodities
-
VNQ – U.S. real estate investment trusts (REITs)
The results are shown in the figure below:
Both metrics demonstrate that the covariance estimates improve as the half-life increases from 5, but they suggest different optimal half-lives. The MVP metric suggests a half-life of around 33 days, while the log-likelihood metric suggests a half-life of around 24 days.
Neither of these metrics is inherently better than the other; they simply measure different aspects of the covariance matrix’s performance. Despite targeting different properties of the covariance matrix, both metrics agree that short half-lives (5-10 days) underperform, and that there is a sweet spot between 20 and 35 days.
The code to replicate this example is:
import yfinance as yf
tickers = yf.Tickers('SPY TLT GLD GSG VNQ')
prices = tickers.download(period='30y', interval='1d')
returns = prices['Close'].pct_change().dropna()
# We'll dump the results of each evalulation
# into a list here.
results: list[dict] = []
# I notice that the resulting curve is neater
# if the x-axis is logged. So we'll use a logspace
# to generate the half-lives.
halflives = np.logspace(
start=np.log10(5), # 5-day half-life
stop=np.log10(100), # 100-day half-life
num=20, # 20 data points
)
for halflife in halflives:
# Shift the returns by one day so that
# the covariance matrix for time t-1 aligns
# with the returns at time t.
covs = (
returns.shift(1)
.ewm(halflife=halflife, min_periods=200)
.cov()
)
results.append({
'halflife': halflife,
'pvar': mvp_metric(covs, returns),
'll': ll_metric(covs, returns),
})
Decoupling
Now that we have a way of measuring covariance estimates independently of backtests, we can explore how to improve the estimates. The first key idea is to decouple estimating variance and correlation.
Note that the correlation between two variables is: $$ \text{corr}(X, Y) = \frac{\sigma^2_{X,Y}}{\sigma_X \cdot \sigma_Y} $$ Where $\sigma^2_{X,Y}$ is the covariance between $X$ and $Y$, $\sigma_X$ is the standard deviation of $X$, and $\sigma_Y$ is the standard deviation of $Y$. We can flip this around to get the covariance in terms of the correlation: $$ \sigma^2_{X,Y} = \text{corr}(X, Y) \cdot \sigma_X \cdot \sigma_Y $$ Which means, if we have estimated the variance (standard deviation squared) of each variable, we only need to estimate the correlation between the two variables to get their covariance.
Now, the reason we want to estimate variance and correlation separately is that they behave differently over time. See the figure below for an example of the variance and correlation of SPY and TLT over time.
Variance lies on the range $[0, \infty)$ and exhibits dramatic spikes followed by a decay back to a mean. Correlation, on the other hand, is bounded between -1 and 1 and tends to be more stable over time, mostly fluctuating around a slowly changing mean.
Ideally, a model should capture the sharp spikes in variance without losing the broader trend in correlation. A short half-life in the EWM helps preserve sudden jumps in variance by avoiding excessive smoothing. Meanwhile, a longer half-life is better suited to capturing the slow-moving changes in correlation over time.
We can estimate variance and correlation independently using a short half-life for variance and a long half-life for correlation. The following code implements this approach:
def ewm_cov(
returns: pd.DataFrame,
var_hl: float,
corr_hl: float,
min_periods: int,
) -> pd.DataFrame:
"""
Returns the exponentially weighted covariance matrix.
Parameters
----------
returns: DataFrame
Asset returns over time.
var_hl: float
Half-life of the variance.
corr_hl: float
Half-life of the correlation.
min_periods: int
Minimum number of periods to consider for the
calculation of the covariance.
"""
vars = returns.ewm(halflife=var_hl, min_periods=min_periods).var()
corrs = returns.ewm(halflife=corr_hl, min_periods=min_periods).corr()
for date in returns.index:
std = np.sqrt(vars.loc[date])
v = np.outer(std, std)
corrs.loc[date] = corrs.loc[date].values * v
return corrs
We’ll conduct an experiment where we use the ETF returns from earlier, vary the half-lives for variance and correlation and measure both the variance of the MVP and the log-likelihood based metric. For the variance estimates, we’ll use the same half-lives as before, for the correlation estimates, we’ll use the half-lives [10, 20, 60]. The results are shown in the figure below.
The figure shows that as the half-life for correlation increases, the performance of the covariance estimates improves. Specifically, across the examined values, a longer half-life for correlation produces better estimates. Under both metrics, the best variance half-life is approximately 18 days, while the best correlation half-life is 60 days.
This example demonstrates the intuition that we should estimate correlation with a longer half-life than variance.
The code to replicate this example is:
from itertools import product
results: list[dict] = []
var_hls = np.logspace(
start=np.log10(5),
stop=np.log10(100),
num=20,
)
corr_hls = [10, 20, 60]
for var_hl, corr_hl in product(var_hls, corr_hls):
covs = ewm_cov(
returns=returns.shift(1),
var_hl=var_hl,
corr_hl=corr_hl,
min_periods=200,
)
results.append({
'var_hl': var_hl,
'corr_hl': corr_hl,
'pvar': mvp_metric(covs, returns),
'll': ll_metric(covs, returns),
})
Shrinkage
A seminal paper came out in 2003 that introduced the idea of shrinkage to estimating a covariance matrix2. The idea is that empirical covariance estimates are often noisy, and this noise can be reduced by blending the estimate with the identity matrix. In other words, we shrink the estimate toward the identity.
Mathematically, we write: $$ \hat{\boldsymbol{\Omega}}_{t-1}^{\text{shrunk}} = (1 - \lambda) \hat{\boldsymbol{\Omega}}_{t-1} + \lambda\boldsymbol{I} $$ Where $\hat{\boldsymbol{\Omega}}_{t-1}^{\text{shrunk}}$ is the shrunk correlation matrix, $\hat{\boldsymbol{\Omega}}_{t-1}$ is the estimated correlation matrix, $\boldsymbol{I}$ is the identity matrix, and $0 \le \lambda \le 1$ is the shrinkage parameter that controls how much we blend the estimate with the identity matrix. When $\lambda = 0$, then no shrinkage is applied.
We can update our ewm_cov function to include shrinkage:
def ewm_cov(
returns: pd.DataFrame,
var_hl: float,
corr_hl: float,
min_periods: int,
shrinkage: float = 0.0,
) -> pd.DataFrame:
"""
Returns the exponentially weighted covariance matrix.
Parameters
----------
returns: DataFrame
Asset returns over time.
var_hl: float
Half-life of the variance.
corr_hl: float
Half-life of the correlation.
min_periods: int
Minimum number of periods to consider for the
calculation of the covariance.
shrinkage: float, default=0.0
Shrinkage factor to apply to the correlation
matrix. A value of 0.0 means no shrinkage,
while a value of 1.0 means complete shrinkage
to the identity matrix.
"""
vars = returns.ewm(halflife=var_hl, min_periods=min_periods).var()
corrs = returns.ewm(halflife=corr_hl, min_periods=min_periods).corr()
for date in returns.index:
std = np.sqrt(vars.loc[date])
v = np.outer(std, std)
corr = corrs.loc[date].values
# Shrinkage
corr = (1 - shrinkage) * corr + shrinkage * np.eye(len(corr))
corrs.loc[date] = corr * v
return corrs
We can run an experience as in the last two sections. We’ll fix the variance half-life to 18 days, and the correlation half-life to 60 days. These are our optimal values from the previous section. We’ll vary the shrinkage parameter between 0.0 and 0.2. We could go up to 1.0 but the curve bottoms out long before 1. The results are in the figure below.
The figure shows that the estimates improve as shrinkage increases from 0.
Conclusion
In this article, we explored three practical ideas for improving covariance matrix estimation. First, we showed how to evaluate estimates without relying on backtests. Second, we saw that variance and correlation behave differently over time and are best modeled separately. Third, we learned about shrinkage, a simple but powerful way to reduce estimation error by blending empirical estimates with a structured target.
Together, these ideas offer a more principled approach to covariance estimation.