Correlation & Covariance: The Math of Moving Together

You already know how to measure a single asset’s risk: take the standard deviation of its returns and call it volatility. One number, one asset, done. But a portfolio isn’t one asset — it’s several, jostling against each other, and the surprise of Modern Portfolio Theory is that how they move together matters more than how wild any one of them is on its own.

Two assets can each be terrifyingly volatile and yet, glued into a portfolio, partly cancel each other out — like two people on a seesaw who, by leaning opposite ways, keep the plank steadier than either could alone. The tool that measures this “moving together” is covariance, and its tidy, unitless cousin is correlation. This lesson defines both from scratch, then shows the exact formula where low correlation pays off as lower risk.

Covariance — how two assets move together

Volatility tells you how far an asset’s returns swing around their own average. Covariance asks a different question about two assets at once: when asset A is above its average, is asset B usually above its average too — or below it?

Think of two dancers. If they’re in sync — both stepping forward together, both back together — they have positive covariance. If they mirror each other — one forward exactly when the other steps back — that’s negative covariance. If they’re oblivious to each other, wandering the floor independently, their covariance is roughly zero.

The precise definition is the average product of each asset’s deviation from its own mean:

$\mathrm{Cov}(R_A, R_B) = \mathbb{E}\big[(R_A - \mu_A)(R_B - \mu_B)\big]$

Read the sign off the product inside. When both assets are above their means in the same period, both deviations are positive, so the product is positive. When both are below, both deviations are negative — and a negative times a negative is still positive. So moving the same direction (up together or down together) pushes covariance positive. Moving opposite directions makes one deviation positive and the other negative, so the product is negative.

• Positive covariance → the assets tend to move the same direction.
• Negative covariance → they tend to move opposite directions.
• Near-zero covariance → no reliable relationship.

A worked example

Here are three years of returns for two assets, A and B:

First the means: $\mu_A = (10 - 2 + 4)/3 = 4\%$ and $\mu_B = (4 + 0 + 2)/3 = 2\%$.

Now the deviations and their products:

Average the products: $\mathrm{Cov}(R_A, R_B) = (12 + 12 + 0)/3 = 8$. The covariance is +8 (percent-squared) — positive, so A and B tend to move together, which matches the table: when A jumped, so did B; when A fell, B fell too.

When it matters

Covariance is the raw ingredient — it’s what actually appears inside the portfolio-variance formula later. But on its own it’s a poor summary, because its size is contaminated by the assets’ volatilities. Use covariance when you’re plugging into a formula; reach for correlation when you want to judge how related two things are.

Correlation — the normalized, unitless version

To fix covariance’s units problem, divide it by the product of the two volatilities. That strips out the scale and leaves a pure, dimensionless number called the correlation coefficient, written $\rho$ (Greek “rho”):

$\rho_{AB} = \frac{\mathrm{Cov}(R_A, R_B)}{\sigma_A\,\sigma_B}$

The magic of that division is that $\rho$ is always trapped in a fixed range: $-1 \le \rho \le +1$. Now the number means the same thing for any pair of assets, so you can finally compare. Read it like this:

• $\rho = +1$ — perfect lockstep. The two move in exact proportion, every time. No diversification possible.
• $\rho = 0$ — unrelated. Knowing one tells you nothing about the other.
• $\rho = -1$ — perfect mirror. When one zigs, the other zags in exact proportion — the ideal hedge.
• Anything in between is a partial relationship, e.g. $\rho = 0.6$ is “fairly correlated but not glued.”

Turning the covariance into a correlation

Take the example above. We found $\mathrm{Cov} = 8$. Suppose the volatilities are $\sigma_A = 6\%$ and $\sigma_B = 2\%$ (computed from the same deviations). Then:

$\rho_{AB} = \frac{8}{6 \times 2} = \frac{8}{12} \approx 0.67$

A correlation of about +0.67 — strongly positive, but not perfect. Notice how much friendlier 0.67 is than “8 percent-squared”: it instantly says “these two move together most of the time, but not in perfect lockstep,” with no units to wrestle and nothing to compare against.

Why it matters for a portfolio

Here’s where it all pays off. Mix weight $w_A$ of asset A and weight $w_B = 1 - w_A$ of asset B. The portfolio’s variance is not the average of the two variances. It’s:

$\sigma_p^2 = w_A^2\sigma_A^2 + w_B^2\sigma_B^2 + 2\,w_A w_B\,\rho\,\sigma_A\sigma_B$

Three terms, and each earns its place:

• $w_A^2\sigma_A^2$ — asset A’s own risk contribution, scaled by how much of it you hold (squared).
• $w_B^2\sigma_B^2$ — the same for asset B.
• $2\,w_A w_B\,\rho\,\sigma_A\sigma_B$ — the cross term, and the star of the show. It’s the only place $\rho$ appears. When $\rho$ is high, this term is big and positive, inflating portfolio variance. As $\rho$ falls, the cross term shrinks; at $\rho = 0$ it disappears entirely; and when $\rho$ is negative it goes negative and subtracts from total variance.

That cross term is where diversification lives. The lower the correlation, the smaller (or more negative) it is, and the more your portfolio’s risk drops below the naive weighted average of the two volatilities. Drag the sliders below and watch the gap open up.

The three cases

To make the cross term concrete, hold a 50/50 mix of asset A ($\sigma_A = 30\%$) and asset B ($\sigma_B = 20\%$), and walk $\rho$ through its three signature values. With $w_A = w_B = 0.5$, the formula becomes $\sigma_p^2 = 0.25(900) + 0.25(400) + 2(0.5)(0.5)\,\rho\,(30)(20) = 225 + 100 + 300\rho = 325 + 300\rho$.

The arithmetic tells the whole story:

• At $\rho = +1$, $\sigma_p$ equals the weighted average, $0.5(30) + 0.5(20) = 25\%$. The assets move in perfect lockstep, so blending them buys you nothing — you just average the risk.
• At $\rho = 0$, the cross term vanishes and risk drops to ≈18.0%, well below the 25% average. Two unrelated assets partly cancel.
• At $\rho = -1$, the cross term goes fully negative and risk collapses to 5% — far below either asset’s own volatility. A perfect mirror lets the swings hedge each other out. (With the right weights you could even drive it to zero.)

Pitfalls

Correlation is powerful and routinely abused. Four traps to keep in mind:

• Correlation is not causation. Two assets can move together because of a shared driver (interest rates, oil prices) or pure coincidence. A high $\rho$ tells you they did move together; it never tells you why, and certainly not that one causes the other.
• ρ is unstable over time — and spikes in crises. A pair that looked uncorrelated for years can snap toward $\rho = +1$ in a crash, when everyone sells everything at once. This is tail dependence: correlations rise exactly when you were counting on them to be low, so diversification evaporates at the worst moment.
• ρ only captures linear co-movement. Correlation measures straight-line relationships. Two assets can be tightly linked in a curved or threshold way and still show $\rho \approx 0$. A low correlation does not prove independence.
• A single ρ hides regime changes. One number averaged over a long window can mask the fact that the relationship was strongly positive in calm markets and strongly negative in stressed ones. The average flatters; the regimes bite.

Key Takeaways

Next up: combining assets into a portfolio and watching the risk–return curve bend — how mixing two assets traces a curve in risk–return space that bows inward exactly because of the correlation math you just learned.