Asset Network Graphs and Minimum Spanning Trees

Take the pairwise correlation matrix of 500 stocks' returns and you get almost 125,000 numbers — technically all the information about how the market co-moves, and practically useless to look at directly. A minimum spanning tree (MST) compresses that firehose into something a human can actually read: it keeps only the $n-1$ strongest connections needed to link all $n$ assets into a single connected network, throwing away every redundant or weaker link.

From correlation to a tree

Correlation isn't quite a distance — two things being perfectly correlated should mean zero distance, and the standard conversion is $d_{ij} = \sqrt{2(1 - \rho_{ij})}$, where $\rho_{ij}$ is the correlation between assets $i$ and $j$. In words: perfectly correlated assets ($\rho=1$) get distance 0, uncorrelated assets ($\rho=0$) get distance $\sqrt{2}$, and perfectly anti-correlated assets ($\rho=-1$) get the maximum distance of 2 — closer co-movement becomes a shorter distance, matching the intuition that similar assets should sit near each other in the network.

With every pair of assets now a distance, an algorithm like Kruskal's or Prim's builds the MST: repeatedly add the shortest remaining edge that doesn't create a cycle, until every asset is connected. The result has exactly $n-1$ edges for $n$ assets — the minimum possible to connect everything — chosen so that the total distance across all edges is as small as it can possibly be.

Worked example

Building an MST over 30 large-cap US stocks using six months of daily returns, three regional banks connect to each other through very short edges ($d \approx 0.35$, meaning $\rho \approx 0.94$), forming a tight cluster; five technology names connect through slightly longer edges ($d \approx 0.55$, $\rho \approx 0.85$); and one gold-mining stock connects to the rest of the tree through the single longest edge in the whole graph ($d \approx 1.25$, $\rho \approx 0.22$) — its only viable link to a market it barely co-moves with. A risk manager reading this tree immediately sees that the gold miner is the position least correlated with everything else, useful information for diversification that would have been buried in a 30-by-30 correlation matrix.

What this means in practice

MSTs are used to visualize sector clustering, to spot genuinely diversifying assets (long, weak edges), and as a first pass before more detailed community-detection algorithms, since the tree's branches are usually a rough preview of the clusters those algorithms will later find formally. An MST recomputed before and after a market stress event is also a standard way to show that correlations tighten in a crisis — the tree visibly compresses, with edge distances shrinking across the board as everything starts moving together.