Direct answer
Correlation Changes refers to the way the measured relationship between two currency returns can vary across time. Timeframe affects this because you are not measuring a single fixed property; you are estimating dependence from a slice of data. Change the observation window (how far back you look) or the holding period (the time span you aggregate into returns), and the estimated correlation can move.
Mechanism and definition
Start with two time series of prices for two currency pairs (or any two tradable instruments). You typically convert prices into returns, then compute correlation on those returns. Correlation answers a narrow question: how consistently the two return series move together within the chosen window, under a specific return definition (for example, simple returns or log returns) and a chosen sampling frequency.
Timeframe enters in two ways:
-
Observation window (estimation timeframe): If you compute correlation using a longer window, you average over more market conditions. If you use a shorter window, you focus on more recent conditions, so the estimate can react quickly to shifts.
-
Holding period (aggregation timeframe): If you measure returns over 1-day intervals, you get one set of return points. If instead you measure returns over 5-day intervals, each data point is an aggregation of several days. Aggregation changes volatility clustering, smoothing, and the timing of shocks—so dependence can look stronger or weaker.
Evidence or example (scenario-impact style)
Consider a scenario where two currency pairs are influenced by mostly the same macro driver during “normal” periods, but during sudden risk-off events they decouple.
- If you estimate correlation using a short window that happens to include multiple risk-off episodes, the estimate can drop (or even change sign) because the returns during those episodes behave differently.
- If you estimate correlation using a long window that mixes both normal and risk-off periods, the estimate may look more stable, but it may also become a compromise—masking the fact that the relationship is regime-dependent.
Now change the holding period. A short holding period can capture immediate co-movement from day-to-day fluctuations. A longer holding period can blend those fluctuations with later adjustments, carry effects, or delayed reactions. The blended returns can therefore show different co-movement than the short-interval returns, even if the underlying drivers are the same.
A key implication is that timeframe changes what you “condition on.” Correlation is conditional on the data segment and return aggregation you choose.
Limitations and risks (material failure modes)
-
Correlation is not a guarantee of future co-movement. Historical correlation reflects observed returns inside a specific window, not a rule that will persist.
-
Regime change and non-stationarity: Currency relationships can shift when market structure changes, volatility spikes, or key risk drivers change. In that case, correlation can be unstable across timeframes.
-
Measurement and methodology choices: Correlation depends on return definition, sampling frequency, missing data handling, and whether you use overlapping vs non-overlapping returns. Different choices can change the estimate.
-
Costs and execution frictions: Even if two series show correlated returns, transaction costs and execution timing can alter realized outcomes. Correlation of returns alone does not account for these effects.
Verification and next question
To independently verify timeframe effects, use a transparent method:
- Choose a consistent return definition.
- Compute rolling correlations using multiple window lengths (for example, short vs long), and compare how the estimate moves.
- Repeat the analysis with different holding periods by aggregating returns over different horizons.
Your control point: if correlation changes materially when you adjust either the observation window or the holding period, that is direct evidence of timeframe sensitivity.
A useful next question is not “what is the current correlation,” but “how stable is the relationship across plausible windows and return horizons, and does it change around known regime shifts?”