Correlation changes: the definition people mix up
Correlation describes how two variables move together over a chosen dataset and time span, typically measured on a scale from -1 to +1. A “correlation change” means the measured relationship differs when you compute correlation over different windows (for example, early vs later periods) or under different assumptions.
A common mistake is treating “correlation changing” as a trading signal by itself. Correlation is descriptive: it summarizes co-movement in the past sample. It does not explain causality, and it does not imply that the relationship will persist.
Another frequent misunderstanding is mixing “currency correlation” with “currency strength” as if they were the same idea. Correlation can shift even when average returns or volatility look stable, and currency strength measures typically rely on different calculations.
Mistake 1: Using correlations as if they were stable laws
What goes wrong: People assume correlation is constant. In reality, correlation often changes because market drivers change (risk-on vs risk-off behavior, liquidity conditions, major news regimes, or shifts in volatility).
Consequence: If you expect prior co-movement to keep working, you can overestimate how reliably two instruments will behave relative to each other.
Neutral check: Clearly state that correlation is window-dependent. If correlation varies materially across adjacent windows, treat that as evidence of non-stability rather than “noise you can ignore.”
Mistake 2: Choosing an inconsistent time window
What goes wrong: Correlations computed over different lookback periods can disagree. A short window may overreact to transient moves; a long window may hide recent shifts.
Example with explicit assumptions: Suppose you compute correlation between two return series using daily returns. If you use one calculation window of 30 days and another of 180 days, you are measuring two different “stories.” If the market regime changed partway through, the longer window averages across regimes.
Consequence: You might conclude the correlation “is positive” or “is negative” without realizing that the sign or magnitude depends on the window.
Neutral check: Test multiple windows (with the same data frequency and method). If conclusions flip, the relationship is not stable under your measurement choices.
Mistake 3: Ignoring the role of the data you feed into correlation
What goes wrong: Correlation depends on what you measure (prices vs returns), how you transform data, and how you handle missing values or non-trading gaps.
Common data pitfalls:
- Using raw price levels instead of returns (price levels can create misleading co-movement).
- Mixing time zones or inconsistent timestamps.
- Changing data frequency mid-analysis (for example, switching from 1-hour to daily without rechecking the method).
Consequence: You may interpret an artifact of preprocessing as a real “correlation change.”
Neutral check: Fix preprocessing rules before comparing windows: same return definition, same frequency, and consistent handling of gaps.
Mistake 4: Treating correlation changes as a standalone indicator
What goes wrong: People often use correlation as if it directly tells them direction or timing. Correlation changes can reflect many things, and the same measured change can coexist with different future outcomes.
Consequence: You can end up with decisions that are not connected to a clear mechanism, making it hard to explain why a particular outcome should occur.
Neutral check: Ask: “What is the decision rule, and does correlation change logically support it?” If correlation is not tied to a mechanism you can describe, it is likely being used as a proxy signal.
Mistake 5: Forgetting costs and execution limits
What goes wrong: Correlation reasoning can overlook practical frictions such as spreads, commissions, slippage, and timing differences.
Consequence: Even if co-movement exists statistically, real execution may reduce or negate the relationship you expected.
Neutral check: When evaluating whether correlation-based ideas are workable, include assumptions about trading costs and execution timing. If the expected effect is small relative to costs, correlation changes may not matter in practice.
Limitations and failure modes to keep in mind
- Non-stationarity: Past correlations may not hold when volatility or market leadership changes.
- Regime shifts: Correlation can move quickly when liquidity and risk appetite change.
- Sampling error: Even with stable underlying behavior, estimates vary by chance, especially over short windows.
- No causality: Correlation does not identify why instruments co-move, so it cannot guarantee what happens next.