Correlation Coefficient

A rolling Pearson correlation between two symbols, scored from -1 to +1 over a chosen lookback window.

What it is

The correlation coefficient measures how closely two series have moved together over a lookback window, on a fixed scale from −1 to +1. A reading near +1 means the two symbols moved up and down together over the window; near −1 means they moved in consistently opposite directions; near 0 means no consistent linear relationship. On TradingView you measure correlation between two symbols by pulling the second symbol's series with request.security() and feeding both series into ta.correlation() with a lookback length — that is the whole recipe, though doing it without introducing repainting takes care (covered below). Traders use rolling correlation to describe relationships such as EURUSD versus GBPUSD, an index versus its constituents, or gold versus real-yield proxies. Two honesty notes. First, correlation is descriptive and historical: it summarizes co-movement in the window you measured, and correlations between instruments demonstrably change over time — a relationship measured today is not a stable property of the pair. Second, correlation is not causation and not a trading edge by itself. This is educational content, not financial advice, and all trading involves risk of loss.

How it works

The underlying statistic is the Pearson correlation: over the last n bars, take the covariance of the two series (the average of the products of their deviations from their respective means) and divide it by the product of their standard deviations. The division normalizes the result into the −1 to +1 range regardless of the instruments' price scales. On each new bar the window slides forward one bar, so the value updates continuously. The implementation detail that matters most on TradingView is how the second symbol's data arrives. request.security() evaluated on the current, still-forming bar returns values that can change until that bar closes — and with lookahead switched on, historical bars can even see data they could not have seen live. Both create repainting: the historical correlation plot looks different from what you would have seen in real time. The correct pattern is to request the other symbol's series offset by one bar (close[1]) with lookahead = barmerge.lookahead_off, and to compare it against your own chart's equally offset series, so both sides of the calculation use only confirmed information. One further subtlety: correlating raw prices of two trending instruments produces inflated readings, because both share a time trend; correlating bar-to-bar changes or returns gives a more honest picture of co-movement.

How traders read it

• Scale: +1 is perfect positive linear co-movement over the window, −1 perfect negative, 0 none. Values are conventionally described as strong beyond roughly ±0.7 and weak inside ±0.3, but these labels are conventions, not thresholds with special meaning.
• It is windowed: a 50-bar correlation describes only the last 50 bars. The same pair can read +0.8 on one window and −0.2 on another — always state the lookback alongside the number.
• Watch the path, not one value: a correlation that has drifted from +0.8 toward 0 tells you the historical relationship has weakened recently; a single snapshot cannot.
• Prices versus returns: correlation computed on raw prices of two trending symbols is usually inflated by the shared trend. Computing it on changes or returns answers the question most people actually mean.
• Diversification context: portfolios built from highly correlated instruments concentrate the same exposure. Note, however, that measured correlations have historically risen sharply during market stress — past low correlation is not a guarantee of future offset.
• A high correlation says nothing about magnitude or cause: two symbols can be 0.9-correlated while one moves ten times as much as the other, and neither may drive the other.

Common settings

Typical lookbacks run from 20 bars (fast, noisy) to 100 or more (slow, smooth); many platform defaults sit at 20 or 50. The comparison symbol and the source series (raw close versus one-bar change or return) matter as much as the length. Reference lines at 0 and around ±0.7 are common. Shorter windows react quickly to relationship changes but jump around; longer windows are statistically steadier but can describe a relationship that has already broken. As with all rolling statistics, changing the window changes every reading, so state your settings when quoting a correlation.

Strengths

• Bounded, standardized output (−1 to +1) that is directly comparable across any pair of symbols and timeframes.
• Textbook, deterministic math — the Pearson formula is fully auditable and implemented natively as ta.correlation() in Pine.
• Makes cross-market relationships visible and quantifiable instead of anecdotal ('these pairs usually move together').
• Useful descriptive input for diversification and hedging discussions, where co-movement is the central question.
• A good teaching vehicle for the request.security() repainting trap, since the naive and correct implementations visibly differ.

Pitfalls to watch

• Repainting via request.security(): requesting the other symbol's current-bar value, or using lookahead_on without a [1] offset, produces a historical plot you could never have computed live. Backtests built on it are invalid.
• Correlation is unstable: measured relationships between instruments drift and sometimes invert. A window that captured a tight relationship says nothing binding about next month.
• Trend inflation: correlating raw prices of two trending series overstates the relationship; use changes or returns for an honest co-movement measure.
• Correlation only captures linear relationships: two series can be strongly related in a nonlinear way and still show near-zero Pearson correlation.
• Outlier sensitivity: a single extreme bar inside the window can swing the coefficient substantially, especially on short lookbacks.
• No predictive power and no causation: correlation describes past co-movement only. It does not forecast either symbol, and no statistic removes the risk of loss.

Pine v6 example

//@version=6
indicator("Rolling Correlation (non-repainting)", overlay = false)

sym    = input.symbol("OANDA:GBPUSD", "Compare Symbol")
length = input.int(50, "Correlation Length", minval = 2)
tf     = input.timeframe("", "Timeframe")

// Confirmed-bar series from the other symbol: [1] offset with
// lookahead_off so historical values match what was visible live.
otherClose = request.security(sym, tf, close[1], lookahead = barmerge.lookahead_off)

// Compare against our own confirmed series so both sides align.
corr = ta.correlation(close[1], otherClose, length)

plot(corr, "Correlation", color = color.blue)
hline(1,    "+1",   color = color.gray)
hline(0.7,  "+0.7", color = color.new(color.gray, 50))
hline(0,    "0",    color = color.gray)
hline(-0.7, "-0.7", color = color.new(color.gray, 50))
hline(-1,   "-1",   color = color.gray)

Pro tip: When quoting a correlation, always attach three things: the window length, whether it was computed on prices or returns, and whether the request.security() call was offset and lookahead-safe — each one can change the number materially, and the third determines whether the historical plot was even achievable live. If a strategy idea depends on a correlation holding, test how that correlation behaved across different windows and regimes before trusting it. Educational content only, not financial advice; relationships between markets change, and all trading carries risk of loss.

Educational only — not financial advice, not a recommendation to trade. No indicator is predictive; trading involves substantial risk of loss.