Covariance tells you the direction of the relationship between two variables: whether they tend to move together or move in opposite directions. A positive covariance means both variables increase (or decrease) at the same time, a negative covariance means one goes up while the other goes down, and a value near zero suggests no consistent linear relationship. The sign is the most important part of interpreting covariance. The raw number itself is harder to make sense of, and understanding why requires a closer look at what covariance actually measures.
What Covariance Actually Calculates
Covariance measures how two variables deviate from their averages in tandem. The calculation works like this: for each pair of data points, you find how far each value sits from its own mean, then multiply those two deviations together. The covariance is the average of all those products.
When both variables are above their means at the same time (or both below at the same time), the product of their deviations is positive. When one is above its mean while the other is below, the product is negative. Covariance sums all of these products and averages them. If the positive products dominate, you get a positive covariance. If the negative products dominate, you get a negative one.
Think of it as a voting system. Each data point “votes” for a positive or negative relationship based on which quadrant it falls in when you center the data at its means. A positive covariance means most of the data points voted for “these variables move together.”
Reading the Sign: Positive, Negative, and Zero
The sign of the covariance is the clearest piece of information it gives you.
- Positive covariance: When one variable is higher than its average, the other tends to be higher too. Example: hours studied and exam scores. More studying generally pairs with higher scores.
- Negative covariance: When one variable is higher than its average, the other tends to be lower. Example: outdoor temperature and heating bills. Warmer months pair with lower energy costs.
- Near-zero covariance: No consistent linear pattern. The variables don’t reliably move together or apart. Knowing one tells you little about the other.
A near-zero covariance doesn’t necessarily mean the variables are unrelated. It means there’s no clear linear trend. Two variables could have a curved or U-shaped relationship and still produce a covariance close to zero, because the positive and negative deviations cancel each other out.
Why the Raw Number Is Hard to Interpret
Here’s the catch: unlike the sign, the magnitude of covariance is difficult to interpret on its own. If you calculate the covariance between height in centimeters and weight in kilograms, you might get a value like 120. Switch height to inches and the covariance changes to something completely different, even though the relationship hasn’t changed at all.
This happens because covariance is expressed in the units of both variables multiplied together. A covariance between height (cm) and weight (kg) is measured in “centimeter-kilograms,” which isn’t intuitive. A covariance of 500 isn’t inherently stronger than a covariance of 50 if the two calculations used different scales or units. You can’t look at a covariance value and say “that’s a strong relationship” the way you might with other statistics. The number depends entirely on the scale of the data.
This is the single biggest limitation when interpreting covariance, and it’s the reason correlation exists.
How Correlation Solves the Scale Problem
Correlation is covariance that’s been standardized. You take the covariance between two variables and divide it by the product of their individual standard deviations. This strips away the units and compresses the result into a scale from -1 to +1.
A correlation of +1 means a perfect positive linear relationship. A correlation of -1 means a perfect negative one. Zero means no linear relationship, just like with covariance. The advantage is that you can now compare the strength of relationships across completely different datasets. A correlation of 0.8 between study hours and test scores is directly comparable to a correlation of 0.8 between rainfall and crop yield, even though those variables have wildly different units.
So if you’re trying to judge how strongly two variables are related, correlation is almost always more useful than raw covariance. Covariance tells you the direction. Correlation tells you the direction and the strength.
Calculating Covariance Step by Step
If you want to compute covariance by hand (or just understand what your software is doing), the process has six steps:
- Step 1: Find the mean of variable X.
- Step 2: Find the mean of variable Y.
- Step 3: For each data point, subtract the mean of X from the X value. These are your X deviations.
- Step 4: For each data point, subtract the mean of Y from the Y value. These are your Y deviations.
- Step 5: Multiply each pair of deviations together.
- Step 6: Average those products. For a population, divide by N (the total number of data points). For a sample, divide by N-1 to correct for the fact that you’re estimating from incomplete data.
A quick example: suppose you have five students and you track their sleep (hours) and quiz scores. If students who slept more also scored higher, the deviation products will be mostly positive, and the covariance will be positive. If better-rested students performed worse (unlikely, but hypothetically), the products would be mostly negative.
Where Covariance Shows Up in Practice
Covariance is a building block in many statistical techniques, even if you don’t always see it directly. In finance, portfolio risk management relies heavily on covariance between asset returns. If two stocks have a high positive covariance, they tend to rise and fall together, which means holding both doesn’t diversify your risk much. If they have negative covariance, losses in one tend to coincide with gains in the other, making the portfolio more stable overall. The Federal Reserve Bank of San Francisco has used covariance matrix forecasts to evaluate risk in currency portfolios, for instance.
In machine learning, covariance matrices help algorithms like principal component analysis identify which features in a dataset carry the most information. In genetics, covariance between traits helps researchers understand which characteristics tend to be inherited together. The concept appears across nearly every field that works with paired measurements.
Key Limitations to Keep in Mind
Beyond the units problem, covariance has a few other blind spots. It only captures linear relationships. If two variables follow a curve, parabola, or some other nonlinear pattern, covariance can understate or completely miss the connection.
Covariance is also sensitive to outliers. A single extreme data point can dramatically shift the result because the deviation products for that point will be large. If your dataset includes unusual values, it’s worth checking whether the covariance is being driven by one or two points rather than the overall trend.
Finally, covariance doesn’t imply causation. Two variables can move together because one causes the other, because they share a common cause, or purely by coincidence. A positive covariance between ice cream sales and drowning incidents doesn’t mean ice cream causes drowning. Both increase during summer months. Covariance describes a pattern in the data, not the reason behind it.