Forex Pair Correlation: Shared Exposure, Returns, and Limits

Pearson correlation measures the direction and strength of a linear relationship between two sets of paired observations. Its coefficient, r, ranges from −1 to +1. A value closer to +1 means the observations tended to move in the same direction; a value closer to −1 means they tended to move in opposite directions. A value near zero indicates little linear association in the chosen sample. It does not rule out a curved or otherwise nonlinear relationship. For example, a symmetric U-shaped relationship can have a Pearson coefficient near zero even when one variable strongly constrains the other. One coefficient cannot summarize every dependency pattern.

The NIST correlation reference expresses the coefficient as the sum of paired deviations from each series’ mean, divided by the product of their standard deviations. Both series must contain matched observations. A coefficient therefore needs context: what was measured, how observations were paired, and which dates were included. The coefficient is standardized and has no pip or account-currency unit: r = 0.8 does not mean an 80% chance of matching moves, an 80% hedge, or an 0.8% price change. There is no universal cutoff that makes a coefficient “high” for every use.

A small invented example makes the arithmetic visible. For three synchronized intervals, let x = [−10, 0, 10] basis points and y = [−10, 10, 0] basis points. Both means are zero; the sum of centered products is 100, and each sum of squared deviations is 200. Pearson’s formula gives r = 100 / √(200 × 200) = 0.5. Three observations are far too few to estimate a stable relationship; this example only illustrates the calculation.

If the question is how two pairs changed together, calculate returns over the same intervals. For a price P, a log return from one observation to the next is ln(Pₜ / Pₜ₋₁); a simple return is Pₜ / Pₜ₋₁ − 1. Choose one convention and use it for both pairs. Matching closing prices by date while using different clock times can pair different market events.

Price levels can trend for long periods, so their correlation may mostly reflect a shared drift rather than short-horizon co-movement. Granger and Newbold’s 1974 paper on spurious regressions90034-7) is a broader time-series warning about misleading relationships among persistent series. For an exposure check, state whether the calculation uses price levels, simple returns, or log returns instead of treating those inputs as interchangeable.

Quote orientation matters too. If one feed reports EUR/USD and another reports the reciprocal USD/EUR, the log returns have opposite signs for the same underlying exchange rate. Use the same pair direction, interval length, timestamp convention, and treatment of missing observations before comparing coefficients.

A long position in a pair buys its base currency and sells its quote currency. That makes the direction of each currency leg visible before looking at a correlation matrix. For a review of base and quote direction, see how to read forex quotes.

Long EUR/USD and long GBP/USD both sell USD, but they buy different base currencies. They share one leg without becoming identical trades. Under synchronized reference quotes, EUR/GBP = EUR/USD ÷ GBP/USD, so its log return is the EUR/USD log return minus the GBP/USD log return. Bid-ask spreads, asynchronous prices, and provider-specific products can make observed quotes depart from that simplified identity.

Correlation alone does not say how much a position can lose. For two P&L series on the same horizon, their combined variance is σ₁² + σ₂² + 2ρσ₁σ₂, where σ₁ and σ₂ are the P&L standard deviations and ρ is their correlation. Position size and volatility enter alongside correlation.

In this calculation, ρ must be the correlation between the two position-level P&L series. A correlation between raw pair returns can stand in only when the positions translate those returns into P&L on the same horizon and currency basis.

Suppose two hypothetical positions each have a one-day P&L standard deviation of $100. At ρ = 0.8, the combined standard deviation is √(100² + 100² + 2 × 0.8 × 100 × 100) = about $189.74. At ρ = 0 it is about $141.42; at ρ = −0.8 it is about $63.25. These are arithmetic examples under fixed assumptions, not loss limits, forecasts, or probabilities of loss. Actual P&L also depends on notionals, quote conversion, execution, and the distribution of returns.

A coefficient is a summary of a particular data set. Daily returns over one year answer a different question from five-minute observations over one day. Mid-prices, bid prices, ask prices, exchange-traded futures, and a retail provider’s rolling product can also produce different series. Include the interval, window, price input, time zone, and provider when recording a value.

OANDA describes a provider-specific tool based on proprietary mid-market prices. Its short-window example extracts synchronized five-minute closing mid-prices and applies Pearson correlation, while its matrix uses a fixed one-year window with New York close data. The page does not describe the example as a return calculation, so do not assume its result matches a matrix built from aligned simple or log returns. It also calls zero “no correlation”; for Pearson, zero means no measured linear association in the sample, not proven independence. This is one provider’s method, not a universal definition or current market-wide estimate.

Window choice also involves a tradeoff: a short window describes recent observations but contains fewer data points, while a long window can blend different policy and volatility regimes. Rolling windows can reveal changes in estimates, but adjacent rolling windows overlap and are not independent confirmations. Session cutoffs and trading frictions affect how a measured series maps to executed P&L; see the forex market-hours guide, spread and commission guide, and slippage and stop-order execution guide.

In a symmetric pair-correlation matrix, the diagonal compares each series with itself and is +1 when the input is nonconstant. Reversing a quote direction can reverse the sign of log returns, so a matrix built from EUR/USD is not interchangeable with one built from USD/EUR. A currency-strength index is another measure with its own construction; it is not the same object as pairwise return correlation.

Currency co-movement is an observed relationship over a chosen period. The BIS Working Paper 727 reports that renminbi and other currency co-movement varied across policy-management phases in its 2015–17 study. The result is a historical research finding about the sample, not a live forecast for a retail pair.

A second BIS study, Working Paper 828, examined reserve composition for 58 central banks over 1999–2017 and found an association between a currency’s co-movement with the dollar or euro and its share in reserves. Official reserves are a different use case from leveraged retail trades. Together, the studies show why currency relationships need a stated sample and institutional context; neither supplies a fixed coefficient for today.

Two return series can be negatively correlated while their positions still have unequal currency notionals, volatility, timing, and costs. A small position in a volatile pair may not offset a larger position in a calmer pair. A hedge assessment needs the account-currency P&L under the same scenarios, plus spread, commission, slippage, financing, and margin effects.

Correlation is usually a linear summary. A Pearson value near zero does not prove statistical independence, and a negative historical coefficient does not guarantee that one position will offset another during a sharp move. If the relationship changes or the positions are mismatched, the portfolio can lose on both legs.

A high coefficient does not show that one pair caused the other to move. Shared currencies, common news, interest-rate expectations, risk appetite, and market structure may contribute to co-movement. The coefficient describes the selected observations; it does not identify a single driver.

Correlation also does not establish cointegration. Correlation summarizes how paired values or returns moved together in a sample. Cointegration asks whether a particular combination of nonstationary price series has a stable long-run relationship under a defined statistical test. A high short-window correlation alone does not establish mean reversion or a pairs-trading opportunity.

Start with the actual positions: pair direction, unit size, entry price, account currency, and product terms. Map which currency each position buys and sells. Then compare synchronized returns using a stated interval and sample window, and record the data source and time of calculation.

For example, a hypothetical long position of 10,000 EUR in EUR/USD at 1.10 has an exposure of +10,000 EUR and −11,000 USD. A separate long position of 10,000 GBP in GBP/USD at 1.27 has +10,000 GBP and −12,700 USD. The combined currency map includes −23,700 USD before any account-currency translation or product-specific adjustment. This is an exposure illustration, not a margin requirement, maximum loss, or proof that a cash-settled product delivers either currency.

Next, estimate each position’s P&L volatility in a common currency and inspect plausible joint moves instead of multiplying a stop-risk percentage by a correlation number. Include costs, overnight financing, margin and close-out rules, and possible execution gaps. The lot-size and pip-value guide explains the conversion step; the margin guide covers account requirements. For pair structure, see major, minor, and exotic pairs, and for interest-rate exposure, see the carry-trade guide.

Recheck the inputs when the time window, market regime, position size, or provider data changes. A matrix can help identify questions to investigate. It cannot replace the position-level cash-flow and risk calculation.