Three open positions: long EURUSD, long GBPUSD, short USDCHF. The platform shows three tickets, three stops and three risk numbers that each look modest on their own. In risk terms it is closer to one position, a short dollar bet at roughly triple size, and a single strong dollar print can take all three stops out inside the same minute.
A correlation matrix exists to make that visible before the loss does. It is a grid of instruments across the top and down the side, with each cell holding a coefficient for how the two moved over a chosen window. A reading near +1 means they tracked each other closely, near 0 means no reliable linear relationship, near -1 means they moved in opposite directions. Decent tools shade the cells so the clusters are obvious without reading a single number.
What the coefficient is actually measuring
Almost every retail tool computes a Pearson correlation on returns over a rolling lookback. Two choices hidden inside that sentence do most of the work.
The first is prices versus returns. Correlating raw price levels of two trending instruments produces high readings for reasons that have nothing to do with shared drivers, because both series drift. Correlating percentage change bar to bar answers the question a trader actually has: when this one moves, does that one move with it?
The second is the window. A 20 day matrix and a 200 day matrix on the same pairs frequently disagree, and neither is wrong. The short window describes the current regime, which is what governs positions you are holding this week. The long window describes the structural relationship, which is what governs whether two instruments belong on the same watchlist at all. If a tool gives you one window and will not tell you which, treat its output as decoration.
Bar size matters as much as lookback. Correlations built on hourly bars capture intraday flow. Correlations on daily bars capture macro positioning. A scalper and a swing trader can look at the same two pairs and honestly reach opposite conclusions, which is why arguments about whether EURUSD and GBPUSD "are correlated" never resolve.
Where you find one
- Charting platforms. TradingView carries a correlation coefficient study that you attach to a chart and point at a second symbol, giving a rolling line rather than a grid. Good for watching a relationship break in real time, weak for scanning twenty instruments at once. Our overview of what TradingView does covers the rest.
- Analytics sites. Myfxbook publishes a correlation page across the majors with selectable timeframes, which is the fastest free grid most traders will find. The Myfxbook guide explains the other tools on that site.
- Terminal indicators. MetaTrader has plenty of community correlation matrices that render a grid in a subwindow from the broker's own history, so the symbols match the ones you actually trade.
- A spreadsheet. Export closes, convert to returns, run a correlation function across the block. Twenty minutes of work and you control the window yourself, with no dependency on someone else's symbol mapping.
Turning the grid into position size
The useful output is a cap on cluster exposure rather than a cap per ticket. If your rule is one percent of the account at risk per idea, and three open trades all sit above 0.7 correlation with each other, the honest reading is that one percent of risk is spread across three tickets rather than three percent across three ideas. Most traders learn this the expensive way, usually on a US inflation release. The arithmetic behind per trade risk sits in our guide to risk management rules.
Negative correlation deserves the same suspicion. Long EURUSD against long USDCHF looks like a hedge on paper and often behaves like one, right up to the point where Swiss franc specific flow arrives and the hedge loses on both legs. Netting off risk because a cell reads -0.8 assumes the relationship survives exactly the event you were worried about, which is when it is least likely to.
A correlation matrix describes the past. It tells you how instruments behaved over the last N bars. Whether they behave the same way through the next central bank decision is a different question, and the grid cannot answer it.
When the grid stops being true
Correlations break, and they break at the worst possible moment. In calm conditions currency pairs trade on local drivers and the matrix looks varied. Under stress almost everything collapses onto one factor, risk on or risk off, and the grid goes uniformly dark. That is precisely when a book built on an assumption of diversification reveals itself as a single directional bet. The behaviour is worth reading about in the context of safe haven flows, because the franc, the yen and gold are the instruments where it shows up first.
There is also a mathematical limit. Pearson correlation measures linear relationships. Two instruments can print a coefficient near zero while having a strong relationship that only appears beyond a certain move size, and the cell will never tell you. Tail behaviour is not in the number.
Then there is the data itself. If a tool computes gold against the dollar index on a different session or a different feed than your broker uses, its cells will not match what your account experiences. That gap is irrelevant for spotting clusters. It becomes relevant the moment someone starts sizing positions to two decimal places off another firm's price history.
A workflow that survives a live account
Look at the grid before you add a position, not after the drawdown. The check takes seconds: find the instrument you are about to trade, read across its row against everything already open, and if anything is above your threshold, either reduce the new size or skip it. That is the whole discipline, and it is the part people abandon first when a setup looks obvious.
Refresh the matrix weekly rather than continuously. Correlations move slowly enough that intraday updates add noise instead of information, and a static weekly reading is easier to apply consistently. Keep a note of the readings you used in your trading journal, because a cluster that blew up will teach you far more when you can see what the grid said beforehand. For the pair by pair relationships themselves, our piece on currency correlations goes deeper into which majors track which.
For a discretionary trader running fewer than five positions, a weekly glance at a free matrix plus a hard cap on same cluster exposure captures most of the available value. Paying for a real time correlation product before you have written down a cluster rule is buying precision you have nothing to do with.
"If two of your open trades share the same driver, you do not have two trades. You have one trade at double size, and the matrix is just the fastest way to see that before the market shows you."
— Alex Onta, Executive Director, SINGUARD
Key Takeaways
- Correlation matrices grade instrument pairs from +1 to -1 over a chosen window, and are best read for clusters rather than exact values.
- Match the lookback and bar size to your holding period, because a 20 day and a 200 day matrix routinely disagree.
- Cap risk per cluster rather than per ticket, so three correlated trades do not quietly triple a single exposure.
- Relationships compress toward one factor under stress, so diversification measured in calm markets can vanish exactly when it is needed.
Frequently Asked Questions
What correlation reading counts as too high?
There is no official threshold. A common working rule is to treat any pair of open positions above roughly 0.7 in absolute value, measured over the window you actually hold trades for, as a single risk cluster and to size the cluster rather than each ticket. The number matters less than having a written rule you apply before adding the position.
Which lookback window should a correlation matrix use?
Match the window to your holding period. An intraday trader gets more from correlations computed on hourly bars over the last few weeks, while a swing trader is better served by daily bars over several months. Looking at both is useful, because a short window that disagrees with the long window is itself a signal that the current regime is unusual.
Do correlation tools work for indices, gold and crypto as well as forex?
Yes, the arithmetic is identical for any two price series. The practical catch is data quality: indices and gold trade on different session hours than spot forex, and crypto trades continuously, so bars do not line up cleanly. Coefficients computed across mismatched sessions are still useful for spotting clusters but should not be used for precise numerical work.