How much risk does a lower correlation actually remove?

Time to put numbers on the mechanism. This is the one formula in the course worth seeing written out, and the payoff is that you will never again have to guess whether a correlation of 0.8 counts as "high."

The two-asset formula

For two holdings with weights w1 and w2, volatilities σ1 and σ2, and correlation ρ:

portfolio variance = w1²σ1² + w2²σ2² + 2 × w1 × w2 × ρ × σ1 × σ2

Then take the square root to get volatility. Three terms: each holding's own contribution, plus a cross-term that carries ρ. All of diversification lives in that third term. When ρ = 1 the cross-term is at maximum, the variance becomes exactly (w1σ1 + w2σ2)², and so the volatility — after the square root — is the plain weighted average. No benefit at all. Every reduction below ρ = 1 shrinks that term, and the portfolio's risk with it.

Worked, three ways

Two holdings, split 50/50, one at 20% volatility and one at 15%. The naive expectation — the weighted average of the two volatilities — is 17.5%. Now the real answer at three correlations:

• ρ = 0.8: variance = 0.010 + 0.005625 + 0.012 = 0.027625 → 16.6%
• ρ = 0.2: variance = 0.010 + 0.005625 + 0.003 = 0.018625 → 13.6%
• ρ = −0.3: variance = 0.010 + 0.005625 − 0.0045 = 0.011125 → 10.5%

Read the column: 16.6%, 13.6%, 10.5% — against a naive 17.5%, and against 15% for the calmer holding on its own. At ρ = 0.2, mixing in a riskier asset produced a portfolio calmer than either piece. That result feels illegal the first time you see it. It is just the cross-term doing its job.

The point about "high" correlations

Notice how little ρ = 0.8 achieved: 16.6% against a naive 17.5%, roughly a 5% improvement. Correlation has to fall a long way before diversification does serious work. This is why "these two are only 80% correlated" is not the reassurance it sounds like — and why hunting for genuinely different exposures is the whole game.

Expected return, meanwhile

Through all three scenarios the portfolio's expected return was identical: the 50/50 weighted average of the two holdings' expected returns. Correlation never appears in the return calculation. Risk moved by six percentage points across the three cases; return did not move at all. Hold that thought — Unit 3 opens with it.

In the data

Running this on two real holdings means lining their daily prices up date by date, and the calendars do not match. Bitcoin and Apple over the same Friday-to-Monday:

Bitcoin trades every day of the week; Apple's market is shut at the weekend, and an exchange holiday removes a day from one series and not the other. Line them up carelessly and you either drop the unmatched days silently or pair one instrument's Monday with the other's Sunday, which moves the correlation you compute without anything moving in the assets. Bitcoin's weekend move has to be counted inside Apple's Friday-to-Monday return.

Try it now

1. Two holdings with clearly different drivers, over the same year. Read a rough yearly swing off each chart, guess ρ from how little the two shapes rhyme, and run the formula. Compare your answer to the naive weighted average.

2. Now stop guessing. Measure the same four quarters on both charts, one quarter at a time, and write the eight percentages in two columns. Count how many quarters share a sign: that fraction is a crude ρ, computed rather than felt, and it is usually further from ±1 than the eye suggests. Then do the whole thing again for two holdings that share a driver, and note how much less the formula rewards you.
3. Say the rule: risk adds in squares with a correlation term, which is why a portfolio can end up calmer than its calmest holding.