What is a Worked Example of Correlation Changes?

Direct answer

A worked example of correlation changes shows how a correlation value can move when the data window, drivers, or conditions used to compute it change. “Worked” means you spell out the numbers and assumptions, then recompute correlation for each situation so you can see what changed and why.

In practice, you treat correlation as a descriptive statistic: it summarizes how two time series move together over a specific sample. If that sample is altered—by using a different period, a different observation frequency, or values influenced by different underlying forces—then the correlation can change.

Mechanism or definition

Correlation (most commonly Pearson correlation) between two variables measures the strength and direction of their linear co-movement in a given set of paired observations. It does not claim the future will match the past.

A “correlation change” therefore refers to a change in the computed correlation when you change something about the setup, such as:

• the time window (earlier vs later periods),
• the event regime (risk-on vs risk-off behavior),
• the sampling frequency or alignment (daily vs hourly, and how timestamps are matched),
• or the inputs (for example, using returns rather than raw levels).

Common technical terms, clarified:

• Time series window: the exact period over which you collect paired observations.
• Paired observations: each point in one series matched with the corresponding point in the other.
• Returns vs levels: returns often behave more stationarily than raw price levels; mixing approaches can alter correlation.

Worked example (numerical, with explicit assumptions)

Assume you have two variables, A and B, measured at five aligned times. You will compute correlation twice using two different samples:

Assumptions (stated explicitly):

1. You use Pearson correlation on the paired observations shown.
2. You compute correlation on the sample of five points each time.
3. Series are already aligned in time (no timestamp mismatch).
4. You use the values directly as given (not returns), purely for demonstration.

Sample 1 (correlation changes are not present yet)

Paired observations (5 times):

• A: 1, 2, 3, 4, 5
• B: 2, 4, 6, 8, 10

Here B is exactly 2×A, so they move perfectly together. Under the Pearson correlation definition, the correlation is +1.

Sample 2 (same variables, different relationship)

Now change only B in a way that flips the co-movement pattern while keeping A the same:

• A: 1, 2, 3, 4, 5
• B: -1, -2, -3, -4, -5

Here B is exactly −1×A, so they move perfectly in opposite directions. The Pearson correlation becomes −1.

What “correlation changed” means here

Between Sample 1 and Sample 2, the correlation moved from +1 to −1. In this toy scenario, the driver behind B’s behavior effectively changed sign. In real data, such sign changes can occur when underlying conditions shift, when a different set of forces dominates, or when the sample window crosses regime boundaries.

Limitations and risks (material failure modes)

1. Correlation is sample-dependent. If you compute correlation on a different time window or with different alignment, you can get a different number. The statistic summarizes the chosen sample, not a guaranteed relationship.

2. Correlation can break under regime shifts. If the relationship between variables is not stable over time (for example, a period dominated by one set of drivers later switches to another), correlation can move sharply.

3. Not causation. Even a strong correlation does not identify why variables move together. Different mechanisms can produce similar correlation.

4. Non-linearity and distribution issues. Pearson correlation targets linear co-movement. If the relationship is non-linear or affected by outliers, the correlation value may be misleading.

5. Small samples can exaggerate changes. With few paired observations, correlation estimates can swing dramatically. Your “worked example” should therefore include enough points to be meaningful.

Verification and next question to ask

To independently verify correlation changes, you can:

1. Fix the definition (which correlation measure, which transformation like returns or levels).
2. Recreate the paired dataset for each window/regime.
3. Recompute correlation for each dataset using the same method.
4. Check sensitivity: does the result remain similar when you slightly adjust the window boundaries or sampling alignment?

A useful next question is: **What exactly changed between samples—time window, alignment, or the underlying drivers?