Direct answer
Pair correlation can look strong, weak, or even opposite depending on the timeframe you use. The timeframe determines how returns are measured, which time periods are included, and whether the calculation mixes short-term behavior with longer-term trends or regime changes.
Mechanism and definition
Pair correlation is a statistical measure of how two time series move together. In a forex context, people typically compute it from a pair of return series, such as returns for currency pair A and returns for currency pair B.
The key variable is the observation window and sampling interval:
- Sampling frequency: whether you use minute, hourly, daily, or weekly data.
- Holding period (aggregation): how long each return step represents (for example, a 1-day return vs a 5-day return).
- Estimation window: how many past return points you include when estimating the correlation.
When you change any of these, you change the underlying data transformation. Shorter intervals emphasize microstructure effects (such as sudden bursts around specific events), while longer intervals can average over them. This is why the same two pairs may show different correlations across timeframes.
A useful way to think about it is time aggregation: if two pairs move together only during certain episodes (for example, risk-on periods), a long timeframe that includes other episodes can reduce the overall correlation.
Example scenario: changing timeframe changes the observed relationship
Assume currency pair A and pair B sometimes move together during a particular market regime, but in other regimes they move independently.
- If you compute returns using a short timeframe (and an estimation window that mostly covers those “together” episodes), the correlation estimate may be high.
- If you compute returns using a longer timeframe that also includes independent regimes, the estimate can drop because the calculation mixes different behavior.
Even if the long-run story is “no consistent co-movement,” you can still observe a temporary correlation on shorter windows. That is not necessarily a contradiction; it reflects that the relationship is not constant across time.
Another practical effect is sampling and overlap. If you change from daily returns to weekly returns, each return step represents a different aggregation of underlying daily movements. Aggregation can turn many small co-movements into a smoother joint movement, or it can cancel them out.
Limitations and failure modes
At least one material limitation is that correlation is conditional on how you measure it. Correlation is not a universal property of two currency pairs; it is an estimate based on a specific timeframe, data frequency, and estimation window.
Common failure modes include:
- Regime mixing: combining periods with different market dynamics leads to correlation estimates that do not represent any single regime.
- Non-stationarity: the statistical relationship may change over time, so past correlation may not match the future.
- Estimation uncertainty: shorter windows produce noisier estimates; longer windows reduce noise but can hide changes.
- Costs and execution: the real-world ability to act on any observed relationship is affected by spreads, commissions, and execution quality. Correlation computed from “raw” returns does not automatically include these frictions.
These limitations mean that correlation can change even when the underlying economic drivers remain similar, because measurement choices alter what the statistic captures.
Verification and next question
To verify your understanding independently, do the following conceptually (without assuming any particular outcome):
- Pick multiple timeframes (for example, short vs long holding-period aggregation).
- Compute the correlation for each timeframe using a clearly stated sampling interval and estimation window.
- Check whether the estimate changes direction or magnitude and whether that coincides with different market episodes.
A good next question is: How sensitive is your correlation estimate to the chosen sampling interval and estimation window length? If the answer is “very,” then the observed relationship is likely timeframe-dependent rather than stable.