Correlation: It Rises When You Need It Low
Correlation is a number between minus one and plus one describing how two series moved together over a chosen window. Plus one is lockstep, zero is no straight-line relation. It measures association, not cause, and it tends to rise in a crisis, exactly when you need it low.
How it works
Correlation is one number describing how two price series moved together over a chosen window. It runs from minus one to plus one: lockstep at plus one, opposite at minus one, and no straight-line relationship at zero.
The scale is bounded, which makes it read like a score. It is not one. It summarises past co-movement across one stretch of bars, and carries none of the context that produced it.
The first limit is that it measures association, not cause. Two markets can move together because one drives the other, or because a third thing drives both. The number looks identical either way.
The second limit is that it only detects linear relationships. Two series can be tightly related in a curved way and still read near zero. Low correlation is not proof of independence.
Why it breaks when it matters
And it rises towards one exactly when you need it low. Assets that behaved independently for years move together in a panic, which is the only period the number was protecting you against.
The third limit is the window. Correlation belongs to the lookback you chose, not to the two markets. Measure over sixty bars and then six hundred, and you get two different answers, both correct.
Which is why a diversified book stops being diversified. Several correlated positions are one position in several accounts, sized as though each were independent. That is how sensible risk per trade still produces an outsized weekly loss.
Two correlated positions are one position paying twice. You pay spread and commission on each leg while carrying a single exposure. The cost genuinely doubles; the risk was never really halved.
In practice
Before adding a position, ask what it shares with what you already hold. The currency, the sector, the direction, the same story. In forex that is usually one currency on three tickets.
The fix is a cap, not a formula. Limit total exposure to a theme rather than only per position, and count correlated trades as one when sizing. That beats any refinement of the coefficient.
One shared shock can create the whole reading. A single opening gap can carry the figure alone, describing one morning rather than a relationship.
And correlated stops all trigger on the same day. Each stop loss looked independent when you placed it, but they watch the same story, so the book closes as one event.
Every round trip costs 2% of a bar. On this site’s shared history that is 0.0098 price units: 2% of a median bar’s range, and 45% of the smallest bar. Doubling an idea doubles the toll.
Correlation and beta measure different things
Correlation and beta are often used as one reading, and they are not. Correlation answers how reliably two series move together, bounded between minus one and plus one, and it says nothing about size. Two markets can be almost perfectly correlated while one moves a tenth as far as the other.
Beta answers the other question: when the benchmark moves one per cent, how far does this instrument typically move? It is unbounded and it carries magnitude, because it combines the strength of the relationship with the ratio of the two standard deviation figures.
Use correlation to decide whether two positions are one idea, and beta to decide how much of the benchmark you have bought. For exchange-traded fund investing and position trading you want both, because a book can be concentrated by relationship and by magnitude at once.
What correlation is not
- Not causation. A shared driver produces the same reading as a direct link.
- Not a measure of size. It says how together two markets move, never how far either one travels.
- Not a constant. It belongs to the window you chose, so quote the lookback with the figure.
- Not a diversification certificate. A low reading in calm conditions promises nothing about a violent week.
When it fails
In a quiet market it reads low and means little. Inside a trading range both series are mostly noise, and noise correlates with nothing. The comforting figure describes the calm.
It fails on curved relationships. Two markets linked through a threshold, where one reacts only once the other has moved far enough, can read close to zero and look independent.
It fails on short windows. A few bars produce a confident figure that is mostly sampling accident, and testing windows until one agrees with you is overfitting by another name.
It fails on long windows too. A longer lookback gives a calmer, staler number that averages away the regime you are actually trading and hides the shift you needed to see.
It fails when one event dominates. A single shock can carry the whole reading, and volume will not warn you, because the coefficient ignores participation entirely.
It fails at portfolio level even when every pair is right. The readings can all look modest while every position leans the same way, which is how a book of separate ideas produces one drawdown.
The original data
A scan of the 31,760 videos in research/search-study-corpus.jsonl, recorded in
research/broker-coverage.json, found 9 with “correlation” in the title: median 1,162 views, 5 channels,
maximum 496,410. “Win rate” appears in 211 videos, median 11,527 views, 130 channels. “Probability” appears
in 83, median 3,118 views, 57 channels. “Overfitting” has 2, median 299 views. “Risk of ruin” and “monte
carlo” have none.
Nine videos on correlation against 211 on win rate. The thing that determines how much you can actually lose in a week is discussed twenty times less than the thing that determines how it feels. That is the finding, and 5 channels against 130 says it is not one creator’s habit.
Then the drawdown figure. In research/series-measurements.json, built by site/measure_series.py over
this site’s shared 576-bar history, 95% of bars sit below a prior peak: deepest 3.76%, median 1.36%, longest
stretch 73 bars, with direction runs averaging 2.01 bars and a longest of 11 across 286 runs. Being
underwater is the ordinary state, and correlated positions turn several ordinary drawdowns into one large
one. Before your next entry, write down the currency, sector and direction it shares with every open
position, then size that group as one trade.
Related
Read risk management next, because correlation only reaches your account through position sizing. Standard deviation explains the dispersion measure correlation is built from, and which beta uses to add magnitude back in. Probability covers the run of losses, and why correlated ones arrive far more often than independent per-trade odds suggest.
I once had four positions open and thought I was spread across four ideas. They were four expressions of one view about the dollar, and they all went wrong on the same afternoon. Nothing in my per-trade rules was broken, which is what made it confusing. I had sized the trades carefully and never sized the group.
— Michael Whitman
This page is educational, not financial advice. Test every idea on your own charts before risking money.