Pearson Correlation in Trading: Diversification, Pairs, and Crisis Risk

Pearson Correlation Coefficient measures the strength and direction of the linear relationship between two asset return series. It tells you whether two assets tend to move together, opposite each other, or independently — and by how much.

Section 1: Core Mechanics

The Pearson Correlation Coefficient (r) measures how consistently the daily returns of two assets move in the same direction. It ranges from -1.0 (perfect inverse relationship — when one rises, the other falls by the same proportion) to +1.0 (perfect positive relationship — they move identically). A value near 0 means no linear relationship exists between the two return series.

Critical rule: Always calculate Pearson Correlation on returns, not on price levels. Price levels are non-stationary (they trend over time), producing spurious high correlation between any two rising assets regardless of any real relationship.

Formula

Where and are the return series of the two assets, and are their standard deviations, and Cov is their covariance. Expanded:

Use daily log returns for both series: .

Inputs

• Series X: Daily log returns of Asset 1 (e.g., GLD)
• Series Y: Daily log returns of Asset 2 (e.g., SPY)
• Rolling window: Number of bars in the correlation calculation (20, 60, or 252 days)

Parameters

Output Range

Visual Behavior

Plot rolling Pearson Correlation as a line panel beneath the price chart. Stable positive correlations appear as a horizontal band in the +0.6 to +0.8 range. A sharp spike toward +1.0 across an entire portfolio indicates a correlation crisis — all assets moving together, diversification failing. A drop toward 0 or below indicates a genuine decorrelation event — potential pairs trade opportunity.

Section 2: Interpretation & Signals

Reading the Correlation Matrix

For a portfolio of N assets, build an N×N matrix where each cell holds the rolling Pearson r between two assets. Diagonal cells are always 1.0 (self-correlation). Off-diagonal cells below ±0.3 represent genuine diversifiers.

Use the correlation matrix to:

1. Identify redundant positions (r > 0.85 — holds two nearly identical assets)
2. Find genuine diversifiers (r < 0.2 — adds true portfolio benefit)
3. Identify hedging candidates (r < -0.5 — moves opposite your core position)

The Crisis Correlation Collapse — The Most Dangerous Assumption in Finance

This is the most important concept in this lesson.

Pearson Correlation estimated during normal bull market conditions severely underestimates how correlated assets become during a market crisis. In March 2020, assets that historically showed correlations of 0.2 to 0.4 with equities — including gold, corporate bonds, REITs, and commodities — all moved toward r approaching 1.0 simultaneously as forced selling dominated every market.

The same pattern appeared in October 2008, August 2015, and March 2022.

Pairs Trading Signal

Two assets with rolling correlation that has dropped from +0.8 to +0.3 may represent a temporary decorrelation — a pairs trade opportunity. But correlation alone is not sufficient for pairs trading — see lesson 46 on cointegration for the proper statistical test.

Chart — Rolling 60-Day Pearson Correlation: GLD vs SPY

Section 3: Pass vs. Live — Real-Time Reliability

Rolling Pearson Correlation updates once per day when all included bars close. There is no repaint risk. The 60-day window means the indicator is always reflecting a 3-month trailing period — useful for portfolio construction but slow to react to sudden regime changes. Use a 20-day window if you need faster reaction to regime shifts.

Section 4: Practical Use Cases

Real example: GLD vs. SPY 60-day Pearson Correlation through 2023 ranged from -0.34 to +0.11, averaging approximately -0.10. This confirmed gold as a genuine portfolio diversifier during the year. However, in March 2020, that same pair showed correlation of +0.48 during the peak selloff week — demonstrating that even gold's diversification benefit partially collapsed under forced selling pressure.

Section 5: Pseudo Code

INPUT: series_x[], series_y[], window=60

PROCESS:
  Step 1: Calculate daily log returns for both series
            ret_x[i] = ln(series_x[i] / series_x[i-1])
            ret_y[i] = ln(series_y[i] / series_y[i-1])

  Step 2: For each bar i from index (window) to end:
            window_x = ret_x[i - window : i]
            window_y = ret_y[i - window : i]

  Step 3: Calculate Pearson r for the window
            mean_x = mean(window_x)
            mean_y = mean(window_y)
            numerator = sum((xi - mean_x) * (yi - mean_y) for xi, yi in zip(window_x, window_y))
            denom_x = sqrt(sum((xi - mean_x)**2 for xi in window_x))
            denom_y = sqrt(sum((yi - mean_y)**2 for yi in window_y))
            r[i] = numerator / (denom_x * denom_y)

  Step 4: (Vectorized alternative)
            import numpy as np
            r[i] = np.corrcoef(window_x, window_y)[0, 1]

OUTPUT: r[] — rolling Pearson r, range -1.0 to +1.0, NaN for first (window) bars
EDGE CASES:
  - If denom_x or denom_y = 0 (zero variance series): return NaN — guard with try/except
  - NaN in either series: propagate NaN for that bar
  - Require minimum 20 observations for statistical validity
  - Always use returns, never raw prices — non-stationary prices produce spurious r near +1.0

Section 6: Parameters & Optimization

Rolling Window Conventions

Parameter Impact

Section 7: Synergies & Conflicts

Section 8: Common Mistakes

Section 9: Cheat Sheet