Diversification means more than owning different things
Owning ten assets doesn't automatically give you diversification. If all ten move the same direction at the same time, you've got concentration risk wearing a diversification costume. What matters is the correlation structure, meaning how much each pair of assets tends to move together. A correlation of +1 means two assets move in perfect lockstep. A correlation of -1 means they move in exactly opposite directions. A correlation of 0 means there's no linear relationship between their moves at all.
The correlation matrix is just a table of the pairwise correlations between every asset you hold. Five assets gives you 10 unique pairs. Twenty assets gives you 190. The count climbs fast, and that's exactly why you want to read the whole matrix instead of eyeballing pairs one at a time. The portfolio behaves like the structure, not like any single relationship inside it.
What typical correlations look like
Inside an asset class, correlations run high. Individual US large-cap stocks correlate with each other around 0.3 to 0.5 on average, though that swings a lot by sector and by regime. US sector ETFs correlate with the S&P 500 at 0.7 to 0.9 in normal markets. International developed equities (EAFE) correlate with US equities somewhere around 0.7 to 0.85, so they give you some diversification but nothing dramatic.
Across asset classes is where it gets useful, because the correlations drop. The stock-bond correlation, measured between the S&P 500 and 10-year Treasuries, stayed negative for most of the 2000 to 2020 stretch, averaging roughly -0.2 to -0.3. That negative number is the whole reason 60/40 portfolios worked as well as they did back then. Stocks fell, bonds rose, and the portfolio's overall volatility got dampened.
Commodities have historically shown low correlations with both stocks and bonds, usually in the 0.0 to 0.3 range. Gold in particular tends to sit near zero against equities over long horizons, though that average hides big swings depending on the environment. Real assets as a group (commodities, TIPS, real estate) correlate more with each other than they do with financial assets like stocks and nominal bonds, which points to a real "real vs financial" split in how returns behave.
Correlations don't hold still
The real problem with correlation matrices is that they're not stable. What you estimated over one period may not survive the next, and they tend to shift in the worst possible way, which is upward during stress. People call it correlation breakdown or contagion, and the effect is that the diversification you measured in calm markets partly evaporates right when you're counting on it.
Stock-bond is the example that costs people the most. From roughly 1960 to 2000 that correlation was mostly positive, so stocks and bonds often moved together. From 2000 to 2021 it flipped mostly negative. Then in 2022 it snapped sharply positive again as stocks and bonds sold off at the same time on rising inflation and rate hikes. A portfolio built around a negative stock-bond correlation behaved nothing like its backtest that year.
Within equities, correlations spike in a sell-off. The average pairwise correlation of S&P 500 stocks can jump from around 0.3 in normal markets to 0.7 or higher during panic selling. That's the "everything went down together" feeling investors already know in their gut. The takeaway is that owning different stocks or different sectors protects you far less in a crash than it does on a quiet Tuesday.
Using the matrix without getting fooled by it
Even with all that instability, the correlation matrix is still the base you build on. The trick is using it with some judgment instead of feeding it straight into an optimizer. One approach is to estimate a separate matrix for each regime (expansion, contraction, crisis) and build a portfolio that holds up across all three rather than one that's tuned perfectly for a single state.
Another is to use longer estimation windows, since a longer window shrugs off recent regime noise, though the cost is it adapts slowly when something genuinely structural changes. A reasonable middle ground is blending a short window (6 to 12 months) with a long one (3 to 5 years).
- Run hierarchical clustering on the matrix to see which assets clump together. Five holdings that all land in one cluster are really one bet, not five.
- Tilt toward assets that form their own cluster, or that sit between clusters, to actually buy yourself diversification instead of redundancy.
- Watch how the clusters shift between your calm-market and stress-market matrices, because that shift is where your protection quietly disappears.
Pairwise correlations miss things
Pairwise numbers don't catch higher-order dependence. Three assets can each show low correlation to the others and still share one underlying risk factor that drags all three down during a specific event. Factor analysis and principal component analysis decompose the matrix into its underlying drivers, so you can see how many genuinely independent sources of return you're holding.
Most diversified portfolios have way fewer independent drivers than they have holdings. Thirty global equity and bond ETFs might come down to 3 to 5 truly independent drivers once you strip out the common factors. That effective dimensionality tells you more than the raw holding count ever will, and it comes straight out of reading the correlation structure carefully. On Blockcircle I lean on this when a portfolio "looks" broad but the matrix says it's really one or two bets, and the fix usually isn't adding more tickers, it's adding a different driver.