What correlation means before it changes
Correlation is a statistical measure of how two price series tend to move together. In practice, it is often estimated from returns (for example, percentage changes over a chosen time interval) rather than from raw price levels.
A correlation “change” means the measured relationship between two currency returns differs across time windows or conditions. This can happen even if the currencies’ long-run drivers are stable, because the market is not stable: the mix of participants, motives, and shocks is constantly changing.
What moves correlation changes?
Correlation shifts usually come from differences in how common factors affect each currency pair, and from time-varying “masks” that temporarily dominate returns.
- Rate and discount-rate expectations Interest rates and expected interest-rate paths influence currency values through relative yield and expected returns. When rate expectations for one currency move more than for the other, the two pairs may start responding differently, reducing correlation.
A simple way to think about this: if both currencies are driven by the same rate surprise, their returns may move together; if only one currency’s rate expectations reprice, co-movement can weaken.
- Macro news and economic surprises Macroeconomic releases (such as growth, inflation, or employment data) can change expectations for policy, risk premia, and future cash flows. If the market interprets new information similarly for two currencies, correlation can rise; if the interpretation differs, correlation can fall.
This is not about the headline itself, but about what the market updates after the headline: expectations, not the printed number, drive returns.
- Risk sentiment and changing risk premia Risk sentiment affects currencies differently. Some currencies are often treated as more sensitive to “risk-on vs risk-off” regimes, while others tend to behave differently when global uncertainty rises or falls. When risk premia move, correlations can temporarily become stronger or weaker depending on which currency is absorbing the stress.
A common failure mode is assuming the same “risk regime” will persist. Correlations estimated during one regime may not hold after the regime changes.
- Liquidity and market microstructure effects Even with identical macro news, correlations can change because liquidity changes how prices form. During calmer periods, trading may be deeper and co-movements may reflect broad valuation factors. During stressed or thin-liquidity periods, price moves can be driven more by positioning, order-flow imbalances, and wider spreads, which can distort estimated correlations.
Liquidity changes do not only affect volatility; they also affect how quickly information is transmitted and how returns co-move.
Evidence or example: why “same shock” can produce different correlation
Consider two currency pairs, Pair A and Pair B, both quoted in different bases. Suppose a macro event leads to a stronger expected policy reaction in one currency than in the other.
- Assumption for the example: both pairs experience the event during the same time window, and returns are computed over equal-length intervals.
- Mechanism: if Pair A’s returns reflect both (1) rate repricing and (2) a broader risk adjustment, while Pair B reflects mainly rate repricing, their co-movement can change when either the rate component dominates or the risk component dominates.
In short, correlations can change because the relative weights of underlying components (rate vs risk vs liquidity) vary over time.
Limitations and risks (what can go wrong when you “use” correlation)
-
Non-stationarity Correlations are not guaranteed to be stable. By construction, any rolling-window or regime-based estimate may show different values because the underlying process changes.
-
Choice of measurement window and return definition Correlation depends on the time interval (minutes vs days), whether you use log or percentage returns, and how you align timestamps. Two analysts can compute different correlations from the same data because the estimation setup differs.
-
Estimation error and noisy inference Short windows produce noisy correlation estimates. Two correlations that look different visually may not be meaningfully different statistically.
-
Hidden common drivers vs causal structure Correlation does not identify causality. Two series can move together because of a third factor (rates, global risk, or liquidity) without either currency “causing” the other.
Verification: how to independently check whether a correlation change is plausible
Use a verification approach that focuses on mechanism rather than prediction: