Diversification and Correlation
Diversification only works to the extent holdings move differently. Counting positions is not the measure; correlation is.
MadStockAlerts Research · Updated August 28, 2026
What to take away
- Twenty correlated holdings behave much like one larger position.
- Diversification reduces company-specific risk and not market risk.
- Correlations tend to rise sharply in a decline, when diversification is most needed.
- Concentration is a legitimate choice, provided it is a choice rather than an accident.
- Hidden correlation runs through factor, supply chain, geography and ownership.
MAD Academy Training Video · 0:45
Diversification Fails When You Need It
Holding many things only reduces risk if they do not move together, and correlations converge toward one in exactly the worst conditions.
This lesson is part of a Stock Alerts + Tools plan.
The measure that matters
Correlation runs from -1 to +1. Two holdings at +1 move identically and provide no diversification whatever. At 0 they are independent. Below 0 they offset, which is what a hedge is.
Ten semiconductor companies is one position expressed ten ways. The account holds ten tickers and one exposure, and the illusion of diversification is the dangerous part.
The illusion is dangerous specifically because it licenses larger total exposure. Somebody who knows they hold one position sizes it accordingly; somebody who believes they hold ten sizes each one as if the others were unrelated.
Two kinds of risk
| Risk | Source | Reduced by diversification? |
|---|---|---|
| Idiosyncratic | A fraud, a failed trial, a lost customer | Yes, substantially |
| Systematic | Rates, recessions, liquidity events | No |
Most of the achievable reduction in idiosyncratic risk arrives within the first twenty or thirty uncorrelated holdings. Beyond that the benefit is small, and what remains is market risk that no number of equity positions removes.
This is the connection to beta. Diversifying reduces the part of volatility that is specific to each company and leaves the part that is shared with the market, which is exactly what beta measures.
Correlations rise in a crisis
The most inconvenient property of correlation is that it is not stable. In a broad decline, assets that behaved independently for years fall together, because the driver is deleveraging rather than anything specific to each holding.
This means diversification is weakest at the exact moment it is most needed, and it is why total exposure matters alongside how that exposure is distributed. A portfolio that is fully invested in thirty uncorrelated names is still fully invested.
Scroll the chart sideways to see all of it.
How many positions actually diversify
The reduction in portfolio variance from adding positions is steep at first and flattens quickly. Going from one position to five removes a large share of the security-specific risk; going from twenty to forty removes very little, because what remains is the risk common to all of them, which no amount of adding removes.
| Positions | Roughly what remains of security-specific risk |
|---|---|
| 1 | All of it |
| 5 | Under half |
| 10 | Around a third |
| 20 | Around a fifth |
| 40 | Around a seventh, and falling slowly |
The figures assume the positions are genuinely unrelated, which is the assumption that usually fails. Twenty positions in one sector are closer to one position than to twenty, because the common factor they share is most of what moves them. Counting positions is therefore a poor measure of diversification, and counting distinct exposures is a better one.
There is also a cost to the far end. Beyond the point where variance stops falling, additional positions add monitoring load and transaction costs while removing almost no risk, and a portfolio too large to follow is exposed to a different failure: not noticing that something has changed.
Correlation is not the same as exposure
Two positions can have a low measured correlation and still be the same bet. Correlation is computed on historical returns over some window, and it is silent about the reason those returns moved together or apart. A homebuilder and a regional bank can show a modest correlation across a calm period and behave identically the moment rates move, because rates were the shared exposure the whole time and nothing had tested it.
| Shared exposure | Where it hides |
|---|---|
| One end market | Suppliers to the same industry, in different sectors on paper |
| Interest rates | Anything with long-duration cash flows, plus anything financed heavily |
| One commodity input | Airlines and chemicals, transport and packaging |
| A single customer | Disclosed in the filings as a concentration, and rarely read |
| The same factor | Several unrelated companies that are all simply cheap, or all growing fast |
The practical reading is that a correlation matrix is a summary of what has already happened and an exposure list is a statement about what would move things together. Both are needed, and the exposure list is the one that survives a regime change, because it was never estimated from a window in the first place.