Direct answer
Common mistakes with pair correlation usually come from confusing “how two price/return series move together” with “a reliable trading relationship.” Pair correlation is a descriptive statistic for historical data. When people treat it as predictive, assume the relationship will persist, or compute it with inconsistent inputs (different timeframes, return definitions, or data cleaning choices), they can reach misleading conclusions.
Mechanism or definition
Pair correlation typically refers to a correlation coefficient computed between two time series—most often returns of two currency pairs (or two derived price series).
Key building blocks:
- Two series must be defined consistently. Decide what you correlate (prices vs returns) and how you compute returns (e.g., simple vs log returns). Correlating raw prices often produces unstable results because prices trend; returns are usually the more interpretable choice.
- A time window must be chosen. The correlation reflects the chosen historical period. A value computed over one window can differ materially from another.
- Correlation is dimensionless and symmetric, but interpretation is limited. A higher magnitude indicates stronger linear co-movement in that period; it does not specify causality, direction for the future, or whether the relationship is stable.
A practical way to think about it: correlation summarizes whether movements in series A tend to coincide with movements in series B within the selected sample.
Evidence or example
Here are common misunderstandings and what goes wrong.
-
Mistake: Treating correlation as a signal. If you compute pair correlation and then conclude that one pair will “follow” the other, you implicitly convert a historical similarity measure into a forecast. Correlation does not encode forward-looking dynamics by itself.
-
Mistake: Mixing inputs (returns definition, sampling frequency, or time alignment). Example assumptions matters: suppose one series uses hourly log returns while the other uses minute simple returns, or the timestamps are not aligned. The correlation result may reflect these inconsistencies rather than any real co-movement.
-
Mistake: Using the wrong data cleaning approach. If missing observations are handled differently across series—such as dropping timestamps in one series but filling gaps in another—then the paired dataset changes. Correlation can shift simply because you changed the sample.
-
Mistake: Over-trusting a single number. A single correlation coefficient may hide changing behavior. For instance, correlation can be moderate overall while being high only during certain market conditions.
-
Mistake: Ignoring the stability question. Even if two pairs were correlated during a prior regime, structural changes (volatility shifts, changes in liquidity, different macro sensitivities) can weaken or invert co-movement.
Limitations and risks
Material limitations and failure modes to keep in mind:
- Correlation is descriptive, not predictive. It summarizes historical co-movement within the chosen window; future relationships can differ.
- Linear-only sensitivity. Standard correlation measures linear association. Two series could have non-linear dependence that correlation does not capture.
- Regime dependence. Correlation often changes when volatility, spreads, or market participation patterns change. This is especially relevant when you later apply the number to a different environment.
- Sensitivity to the window and outliers. Short windows can produce unstable estimates; a few extreme observations can disproportionately influence the result.
- Costs and execution are not included. Correlation does not consider transaction costs, slippage, or execution timing. Any decision based on correlation alone ignores these real-world frictions.
Verification or next question
To verify pair correlation claims neutrally, use a “consistency and stress” checklist:
- Recompute with consistent inputs. Use the same return definition, sampling frequency, and aligned timestamps for both series.
- Change the window and compare stability. Compute correlation across multiple windows (e.g., earlier vs later samples) to see whether the estimate is persistent or fragile.
- Check robustness to cleaning. Apply a consistent missing-data rule for both series and note how much correlation changes.
- Separate regimes conceptually. If correlation changes when volatility changes, treat that as information about regime dependence rather than a stable relationship.
- Ask what the number does not answer. Correlation does not establish causality, does not guarantee continuity, and does not confirm future direction.
A useful next question: “What exact series, transformations, and window definition produced this correlation value?” If you cannot state those choices precisely, you cannot independently verify the result.