Direct answer
Pair correlation is an historical relationship between two currency pairs: it summarizes whether their returns tend to rise or fall together, and how strongly. You can interpret it as a descriptive measure, not as proof of cause, not as a stable property, and not as a guarantee of how prices will move in the future.
A useful goal is to translate correlation into expectations about co-movement only for the specific period and calculation setup used. Then verify whether that expectation still holds by recomputing the same correlation on new, out-of-sample data.
Mechanism and definition
Pair correlation typically refers to the Pearson correlation coefficient computed from two return series. In plain terms:
- Pick a start-to-end price series for each currency pair.
- Convert prices to returns (for example, simple returns or log returns).
- Choose a time window (for example, daily returns over 90 days, or intraday returns over a specific range).
- Compute the correlation between the two return sequences.
Interpretation is relative to that setup:
- A positive value means the pairs’ returns often move in the same direction during the sampled period.
- A negative value means they often move in opposite directions.
- A value near zero means there is little linear co-movement in that sample.
Important: correlation describes the co-variation of returns, not the level of prices, and not whether one pair “drives” the other.
Evidence or example (with explicit assumptions)
Assume you compute daily returns for Pair A and Pair B over one 60-day period, using the same return formula for both. If the computed correlation is +0.80, that indicates strong historical tendency for their daily returns to have the same sign more often than chance would suggest, within that 60-day period.
Now change only one assumption: use a different 60-day window immediately after the first period. It is possible to observe a much lower correlation even if both pairs remain “the same instruments.” This can happen because market relationships can shift when trading conditions, volatility patterns, or macro-driven flows change.
Also consider a failure mode: if you inadvertently mix data frequencies (for example, Pair A computed from daily closes and Pair B computed from intraday snapshots), the correlation you get may reflect data handling differences rather than real co-movement.
Limitations and risks
At least four material limitations affect how you should interpret pair correlation:
-
Time-window dependence Correlation is not a fixed constant of the pair. A different sample period can produce a different correlation value.
-
Model and return-definition dependence Different choices (simple vs log returns, different sampling frequency, different roll or alignment rules) can change the computed correlation. Correlation is sensitive to how the numbers are constructed.
-
Linear dependence only Pearson correlation captures linear co-movement. Two series can have a relationship that is nonlinear or changes shape across volatility regimes, yet still show a low or unstable Pearson correlation.
-
Not a causal or predictive tool Even if correlation is stable historically, it does not establish causality. Historical relationships can weaken or reverse, especially around regime changes.
-
Costs and execution effects Correlation computed from mid prices or closes does not automatically include trading costs, spreads, slippage, or execution timing. Two pairs can appear correlated on paper while differing in realized outcomes.
Verification and next question
To verify the interpretation independently:
- Recompute the correlation using the same method but multiple, non-overlapping time windows.
- Check whether correlation remains similar across windows or collapses after a regime shift.
- Ensure your return series are aligned in time and computed with the same formula and frequency.
A good next question is: What correlation result should count as “stable” for your chosen purpose? The answer depends on your tolerance for change and on how you defined returns and windows.