What is correlation actually measuring?
Correlation is the most load-bearing number in portfolio theory, and it is routinely misread. A few minutes on what it does and does not say will save you from most of the classic mistakes.
The definition, in plain words
Correlation measures how consistently two things move in the same direction, on a scale from +1 to −1.
- +1 — every time one rises, so does the other. Perfect synchrony.
- 0 — knowing one's move tells you nothing about the other's.
- −1 — perfectly opposed: one zigs precisely when the other zags.
Real financial correlations almost never sit at the extremes. Two large banks in the same country might run around +0.8. A US equity index and a European one, often somewhere in the +0.7 to +0.9 region. Stocks and gold have historically wandered around zero, drifting either side. Those are illustrations of typical ranges, not constants — a point Unit 4 will hammer.
Three things correlation does NOT tell you
It says nothing about magnitude. A stock that moves 0.1% every time the market moves 1% can still have a correlation of +0.9 with it. Correlation captures direction and consistency, not size. Size is a different measure (beta), and confusing the two is the most common error in this subject.
It says nothing about causation. Two assets can be correlated because one drives the other, because a third thing drives both, or by pure coincidence over a short sample. Correlation describes the data; it does not explain it.
It is not a permanent property. Correlation is estimated from a specific window of history. Change the window — daily instead of monthly, 2019 instead of 2022 — and the number changes, sometimes dramatically.
Measured on returns, not prices
One technical point that matters: correlation in portfolio work is computed on returns (percentage changes), not on price levels. Two stocks that both drifted upward for a decade will show a high correlation of prices while their day-to-day returns may be nearly unrelated. Eyeballing two price charts is a rough version of the right calculation — useful for intuition, but it flatters correlation upward, because everything that trends looks alike.
A worked reading
Suppose that over one year an airline rose on 130 days and fell on 120, while an energy producer rose on 128 and fell on 122. Similar-looking totals. But if the airline's up-days are mostly the energy producer's down-days — cheap fuel helping one and hurting the other — the correlation is negative even though both charts end near where they started. The pattern of pairing, not the pair of outcomes, is the signal.
In the data
The returns this calculation needs come from adjusted prices, which carry splits and dividends. Apple across its 2020 split shows why:
On the traded price Apple fell three quarters in a day; on the adjusted price it rose about 3%. Take percentage changes off the traded price and every split and every dividend enters the series as a one-day collapse — a manufactured move that lines up with nothing in the other series, which drags the measured correlation toward zero.
Try it now
- Two instruments with historically different behaviour are below, over their full histories. Navigate both to a calm year — 2017 does nicely — and count, week by week, how often the two lines lean the same way. You are computing a correlation with your eyes, which is enough to see what the number is made of.
- Now move both to 2020 and do it again. Same two instruments, different window, different answer. That is the third caution above, measured rather than asserted — and Measure each chart across each window if you want the two numbers rather than the impression.
- Say the caution aloud: correlation is direction and consistency over a chosen window — not size, not cause, and not a constant.