Correlation and covariance are the two statistical concepts that make diversification mathematically possible. They describe how two assets’ returns move in relation to one another — and that relationship, more than any single asset’s individual risk or return, determines how much a portfolio’s overall risk can actually be reduced by combining assets together.
This guide explains covariance and correlation from the ground up — their formulas, how they relate to one another, worked examples, how they drive portfolio-level risk reduction, the correlation matrix used in multi-asset portfolio construction, rolling correlation, and the real-world limitations that every investor relying on historical correlation data should understand.
Key Takeaways
- Covariance measures the direction in which two assets’ returns move together, while correlation standardizes that relationship to a scale between -1 and +1.
- Correlation is calculated by dividing the covariance between two assets by the product of their individual standard deviations.
- Lower correlation between portfolio holdings generally allows for greater risk reduction through diversification, holding individual asset risk constant.
- A correlation matrix extends this concept across many assets simultaneously, and is a required input for full portfolio optimization.
- Rolling correlation reveals how the relationship between two assets changes over time, rather than relying on a single static historical figure.
- Correlations tend to rise sharply during market crises, often just when diversification is needed most — a well-documented and important limitation.
- Correlation measures linear relationships specifically; two assets can have a strong non-linear relationship while showing low correlation under the standard measure.
Covariance: Measuring the Direction of Co-Movement
Covariance measures whether two assets’ returns tend to move in the same direction or opposite directions, and roughly how strongly. Its formula, for two assets over a set of historical return observations, is:
Covariance(A,B) = Σ[(Return of A − Average Return of A) × (Return of B − Average Return of B)] / (n − 1)
Interpreting the Sign of Covariance
- Positive covariance: The two assets’ returns tend to move in the same direction — when one is above its average, the other tends to also be above its average.
- Negative covariance: The two assets’ returns tend to move in opposite directions — when one is above its average, the other tends to be below its average.
- Covariance near zero: There’s little discernible directional relationship between the two assets’ movements relative to their respective averages.
The Problem With Covariance Alone
Covariance’s magnitude is difficult to interpret on its own, since it’s expressed in units that depend on the scale of the two assets’ returns — a covariance figure of 0.003 might represent a very strong relationship for two low-volatility assets, but a comparatively weak one for two high-volatility assets. This scaling problem is exactly what correlation is designed to solve.
Correlation: Standardizing the Relationship
Correlation takes covariance and standardizes it into a consistent, bounded scale by dividing by the product of the two assets’ individual standard deviations:
Correlation(A,B) = Covariance(A,B) / (Standard Deviation of A × Standard Deviation of B)
This produces the correlation coefficient, always bounded between -1 and +1, which makes it directly comparable across any pair of assets, regardless of how volatile each individual asset happens to be:
- Correlation of +1: Perfect positive relationship — the two assets always move in the same direction, in exact proportion.
- Correlation of 0: No linear relationship between the two assets’ movements.
- Correlation of -1: Perfect negative relationship — the two assets always move in exactly opposite directions, in exact proportion.
- Correlation between 0 and +1: A positive, but imperfect, tendency to move in the same direction — the most common real-world case for assets within similar market exposures.
- Correlation between 0 and -1: A negative, but imperfect, tendency to move in opposite directions.
A Worked Example
Consider two hypothetical assets with the following simplified quarterly returns over four periods:
- Asset A: 4%, 2%, -1%, 3% (average = 2%)
- Asset B: 3%, 1%, -2%, 2% (average = 1%)
Both assets moved in the same direction relative to their own averages in every single period — above average together, below average together — producing a strongly positive covariance, and, once standardized, a correlation coefficient close to +1 (in this simplified, illustrative case). If Asset B had instead moved in the opposite direction from Asset A in each period, the same calculation would produce a strongly negative correlation instead, close to -1.
This simplified example illustrates the core mechanic: correlation isn’t about whether two assets have similar average returns or similar volatility — it’s specifically about whether their period-to-period movements, relative to their own respective averages, tend to align.
Why Correlation Drives Diversification
The mathematical link between correlation and portfolio risk reduction comes directly from the portfolio variance formula for a two-asset portfolio:
Portfolio Variance = w1²σ1² + w2²σ2² + 2(w1)(w2)(σ1)(σ2)(ρ)
Where w1 and w2 are portfolio weights, σ1 and σ2 are each asset’s standard deviation, and ρ (the Greek letter rho) is the correlation coefficient between the two assets. The correlation term directly scales the final piece of the formula: as correlation falls from +1 toward 0 and beyond toward -1, that final term shrinks, and eventually turns negative, pulling total portfolio variance down below what a simple weighted average of the two assets’ individual variances would produce.
This is the entire mathematical basis for the well-known claim that “diversification is the only free lunch in investing” — lower correlation between portfolio holdings allows an investor to reduce total portfolio risk without necessarily sacrificing expected return, since expected portfolio return is a simple weighted average unaffected by correlation, while portfolio risk is not.
The Correlation Matrix
Building a Matrix Across Many Assets
When a portfolio includes more than two assets, the pairwise correlation between every possible combination of assets needs to be considered, not just each asset’s relationship to the portfolio as a whole. This full set of pairwise correlations is organized into a correlation matrix — a grid showing the correlation coefficient between every pair of assets in the portfolio’s investment universe.
How Quickly the Matrix Grows
The number of unique pairwise correlations required grows quickly as more assets are added, following the formula n(n-1)/2 for n assets. A 10-asset universe requires 45 unique correlation estimates; a 20-asset universe requires 190; a 50-asset universe requires 1,225. This rapid growth is a significant practical challenge in constructing reliable correlation matrices for larger, more diversified portfolios, and is one of the key reasons full mean-variance optimization becomes progressively more difficult, and more sensitive to estimation error, as the number of assets under consideration increases.
Using the Matrix in Portfolio Construction
The full correlation matrix, combined with each asset’s expected return and standard deviation, is precisely the input set required to calculate the efficient frontier and perform mean-variance portfolio optimization across a multi-asset universe, as discussed in more detail in dedicated coverage of the efficient frontier and portfolio optimization.
Rolling Correlation: Correlation Changes Over Time
A single, static correlation figure calculated over an entire historical period can obscure meaningful changes in the underlying relationship over time. Rolling correlation addresses this by calculating correlation over a moving window — for example, the trailing 60 trading days — and updating that figure as the window moves forward through time, producing a correlation time series rather than a single fixed number.
Plotting rolling correlation between two assets often reveals that their relationship isn’t constant — two assets might show low or even negative correlation during calm market periods, then show sharply rising correlation during periods of market stress, a pattern that has real consequences for diversification, discussed further below.
Why Rolling Correlation Matters for Risk Management
An investor relying on a single long-term average correlation figure to estimate diversification benefits may be significantly underestimating how correlated their holdings could become during precisely the market conditions when diversification matters most. Monitoring rolling correlation provides a more dynamic, current view of how a portfolio’s diversification characteristics are actually behaving, rather than assuming a static historical average will hold indefinitely.
Correlation Breakdown During Market Crises
One of the most consequential and well-documented patterns in correlation data is that correlations across many assets — including assets that appeared to have low or negative correlation during calm periods — tend to rise sharply during periods of broad market stress or crisis. This phenomenon is sometimes summarized with the observation that “in a crisis, all correlations go to one.”
Why This Happens
During a genuine market crisis, broad risk-aversion can drive simultaneous selling across many different asset classes at once, as investors and institutions seek liquidity or reduce risk exposure across their entire portfolio simultaneously, rather than in an asset-specific way. This broad-based, correlated selling pressure can override the more idiosyncratic, asset-specific factors that produced lower correlation during calmer periods.
The Consequence for Diversification
This pattern means that a portfolio constructed to be well-diversified based on historical average correlations can behave quite differently — and offer less protection — during precisely the periods when an investor most needs that diversification to work as expected. This is the same underlying dynamic discussed as a core criticism of Modern Portfolio Theory more broadly, and it’s a significant reason many risk management frameworks incorporate stress testing and scenario analysis alongside standard correlation-based portfolio construction.
Correlation Measures Linear Relationships Only
The standard correlation coefficient specifically measures the strength of a linear relationship between two variables. Two assets can have a strong, genuine, but non-linear relationship — for example, one asset’s returns might depend on another’s in a more complex, curved pattern — while still showing a low standard correlation coefficient, since that coefficient isn’t designed to capture non-linear dependencies.
This is a more technical limitation than the crisis-correlation issue, but it’s worth understanding: a low correlation coefficient means “no strong linear relationship detected,” not necessarily “no meaningful relationship exists at all.” More advanced dependency measures exist specifically to address this limitation, though standard correlation remains the dominant, most widely used measure in practical portfolio construction due to its simplicity and direct mathematical link to portfolio variance.
Practical Applications in Portfolio Construction
- Asset allocation across asset classes: Combining asset classes with historically lower correlation to one another — such as domestic equities, international equities, bonds, and alternative investments — is a foundational diversification strategy built directly on correlation.
- Evaluating new additions to a portfolio: Before adding a new holding, examining its historical correlation to existing holdings can reveal whether it’s likely to provide genuine diversification benefit, or whether it largely duplicates risk already present in the portfolio.
- Constructing the efficient frontier: The full correlation matrix, alongside expected returns and volatilities, is a required input for mean-variance portfolio optimization.
- Stress testing diversification assumptions: Examining how correlations behaved during specific past crisis periods, rather than relying only on long-run average correlations, provides a more realistic picture of how a portfolio might behave during future stress.
- Monitoring portfolio risk over time: Tracking rolling correlation between key holdings can serve as an early indicator of shifting portfolio-level risk characteristics, even before overall volatility measures change meaningfully.
Correlation vs Covariance: Quick Reference
| Feature | Covariance | Correlation |
|---|---|---|
| Scale | Unbounded; depends on units of the underlying assets | Always bounded between -1 and +1 |
| Directly comparable across asset pairs? | No — magnitude depends on each asset’s volatility | Yes — standardized scale allows direct comparison |
| Used directly in portfolio variance formula? | Yes, in an alternative form of the formula | Yes, in the most commonly used form of the formula |
| Intuitive interpretation | Harder — requires context on each asset’s scale | Easier — a single number from -1 to +1 with clear meaning |
Frequently Asked Questions About Correlation and Covariance
What is the difference between correlation and covariance?
Covariance measures the direction in which two assets’ returns move together but is expressed in units that depend on the assets’ scale, while correlation standardizes covariance into a bounded scale from -1 to +1, making it directly comparable across any pair of assets.
How does correlation affect portfolio diversification?
Lower correlation between portfolio holdings allows for greater reduction in total portfolio risk when combined, since the correlation term directly scales the covariance component of the portfolio variance formula, with correlation below +1 reducing risk below a simple weighted average.
What is a correlation matrix?
A correlation matrix is a grid showing the correlation coefficient between every pair of assets in a portfolio’s investment universe, required as an input for mean-variance optimization and constructing the efficient frontier across more than two assets.
What is rolling correlation?
Rolling correlation calculates correlation between two assets over a moving historical window, such as the trailing 60 trading days, producing a correlation time series that reveals how the relationship between the assets changes over time, rather than relying on a single static average.
Why do correlations rise during market crises?
During market crises, broad risk-aversion often drives simultaneous selling across many different asset classes as investors seek liquidity or reduce risk exposure across their entire portfolio at once, which can override the more idiosyncratic factors that produce lower correlation during calmer periods.
Does low correlation mean two assets have no relationship at all?
Not necessarily. The standard correlation coefficient measures linear relationships specifically, so two assets could have a genuine but non-linear relationship while still showing a low standard correlation coefficient.
Can correlation be negative in practice?
Yes. Some asset pairs have historically shown negative correlation, meaning they tend to move in opposite directions, though negative correlation between major asset classes is less common and less consistent than positive or near-zero correlation, and can change over time.
Final Thoughts
Covariance and correlation aren’t just abstract statistics — they are the specific mathematical mechanism through which diversification actually reduces portfolio risk. Understanding how they’re calculated, how they scale into a full correlation matrix for multi-asset portfolios, and importantly, how they tend to shift — often unfavorably — during periods of market stress, is essential for building portfolio construction decisions on a realistic, rather than overly optimistic, foundation.
Diversification isn’t just about holding many different assets. It’s about holding assets whose correlation is genuinely low — and understanding that this correlation isn’t fixed, especially not during the moments when a diversified portfolio is being counted on most.
