I opened a book once that had twelve names in it and felt pretty good about that. Twelve is a lot. Then I ran the correlation matrix and most of those names moved together on the same handful of days. When the tape turned, they all turned. What looked like twelve bets was really closer to three, and the other nine were mostly there to make the position count look respectable.
Position count is the number everyone quotes because it is the easiest one to see. It tells you almost nothing about how concentrated your risk actually is. Two positions in perfectly uncorrelated things are more diversified than twenty in things that all rhyme. So the question I keep coming back to is whether there is a single number that captures the real spread, one you can watch over time and defend to someone who asks why the book bled out all at once.
Why the count lies to you
Diversification is about how independently your positions move, not how many of them there are. If everything in the book responds to the same driver, you own one bet in twelve costumes. The costumes cost you fees and attention and they give you nothing back when it matters, because they all fail together on exactly the day you needed them not to.
This shows up constantly in crypto because so much of the market is a leveraged bet on the same beta. You can hold a dozen tokens across a dozen chains and think you have spread your risk, and then one bad macro print takes the whole thing down twenty percent in a session. The names were different. The bet was the same. The correlation matrix would have told you that before the drawdown did, if you had bothered to look at it.
Turning the matrix into one number
There are two related ways to squeeze a correlation or covariance matrix down to a single honest number. Both are worth knowing because they answer slightly different questions.
The first is the diversification ratio. You take the weighted average of your individual position volatilities, then divide it by the actual volatility of the whole portfolio. If your positions were perfectly uncorrelated, the portfolio vol would be much lower than the average of the parts, so the ratio would be high. If everything is correlated at one, the ratio collapses toward one, meaning the portfolio behaves like a single leveraged position. A ratio near one is the number screaming at you that your names are redundant.
The second is the effective number of bets, and this is the one I actually track because it is intuitive. The clean version comes from principal component analysis. You take your covariance matrix and decompose it into independent factors, each with its own share of the total variance. Then you treat those variance shares like a probability distribution and compute what is called the entropy of that distribution, and exponentiate it. The output is a number that reads like a position count but tells the truth. If your twelve names load onto three underlying factors that explain almost all the variance, the effective number of bets comes back around three, not twelve.
If PCA feels like too much machinery, there is a rougher proxy that gets you most of the way. Compute the average pairwise correlation across your book. As that average climbs toward one, your effective bets shrink toward one regardless of how many lines you hold. It is not as precise as the entropy version but it moves in the right direction and you can compute it in a spreadsheet.
A workflow you can actually run
Here is roughly how I turn this from a concept into something on a dashboard.
- Pull a trailing window of returns for every position, typically ninety to a hundred and eighty days of daily data. Shorter windows are jumpy, longer windows smear over regime changes, so pick one and keep it consistent.
- Build the correlation matrix and, if you can, the covariance matrix weighted by your actual position sizes. Sizes matter. A tiny uncorrelated position does not rescue a book dominated by three correlated whales.
- Compute the diversification ratio and the effective number of bets. Write both down next to your position count.
- Track the gap between position count and effective bets over time. When that gap widens, your book is getting quietly more concentrated even though you have not added or removed a single name.
The rule of thumb I use is simple. If my effective number of bets is less than half my position count, I am fooling myself about diversification and I need to either trim the redundant names or add something that genuinely moves on a different driver. Something market-neutral, or a bet keyed to a different catalyst entirely, like an insider-flow or political-flow signal that does not care what beta did that morning.
Where this quietly breaks
A few failure modes worth naming, because I have walked into all of them. Correlations are not stable. The comfortable low-correlation number you measured in calm months compresses toward one in a crisis, which is precisely when you were counting on the diversification. So it is worth computing your effective bets on a stressed window, not only the calm one, and planning around the smaller number.
Short lookbacks lie in the other direction too. Two things can look uncorrelated for ninety days purely by luck and then reveal they share a driver the moment that driver wakes up. And the whole calculation assumes your return series are clean. A stale price or a thinly traded name will understate its true correlation and flatter your effective bets, so garbage in still gets you garbage out here.
None of this replaces judgment about what you own and why. But when someone asks how diversified the book is, I would rather hand them a number that survives a look at the correlation matrix than point at a list of twelve tickers and hope. The effective number of bets is that number. It is not hard to compute and it is annoyingly honest, which is exactly what you want from a risk metric.