Cluster analysis · ALB Equity
Sector limits are how most mandates describe risk. Grouped instead by what the model actually measures, 943 companies fall into 6 profiles that cut straight across the 9 GICS sectors — and the sector label turns out to account for only 12% of what makes these companies alike.
Release 2026.06 · 30 June 2026 · recomputed from the model output on every release, not maintained by hand.
The finding
Each bar is the share of variation in the model’s own readings that a grouping accounts for. The comparison only means something with all three present — see the note beneath it.
The classification most mandates and risk reports are written in.
Far more groups, hand-built — the fairest test of classification alone.
Fitted to maximise exactly this. Shown for scale, not as a fair contest.
Sector limits are the instrument most portfolios use to control concentration. If the sector label captures a modest share of what makes companies behave alike, a book can satisfy every sector limit it has and still be a single bet — long cash generation and short momentum, say, with nothing in the sector report to show it.
The point is not that sectors should be dropped. They are how mandates, benchmarks and reporting are written, and that will not change. It is that a second lens is needed alongside them, and that the second lens should come from the data rather than from another committee.
The profiles
Named from their own two strongest analytical families — generated from the data, not written by an analyst, so they stay accurate when the next release moves the boundaries.
Most represented
Most represented
Most represented
Most represented
Most represented
Most represented
Profile detail
| Analytical family | C1n=206 | C2n=182 | C3n=173 | C4n=141 | C5n=130 | C6n=111 |
|---|---|---|---|---|---|---|
| Growth | -0.69 | -0.27 | +0.84 | +0.45 | -0.93 | +0.92 |
| Free cash flow yield | -0.45 | +0.96 | +0.26 | -0.75 | +0.96 | -1.32 |
| Expectations & implied growth | -0.81 | +0.68 | +0.72 | -0.21 | +0.57 | -1.12 |
| Traditional valuation | +0.13 | +0.90 | -0.32 | -1.15 | +0.75 | -0.64 |
| Margins & profitability | -0.12 | -0.32 | +0.67 | +0.27 | +0.06 | -0.71 |
| Leverage & management behaviour | -0.74 | -0.25 | +0.67 | +0.76 | -0.25 | +0.07 |
| Market indicators | +0.16 | -0.42 | -0.38 | +0.58 | +0.36 | -0.20 |
| Quality | -0.78 | -0.09 | +0.88 | +0.58 | +0.45 | -1.04 |
| Relative performance | +0.36 | -0.81 | -0.67 | +0.93 | +0.49 | -0.05 |
Standard deviations from the universe mean, on the model’s own analytical families. Positive is the favourable direction throughout — the engine scores 0 as best, and these readings are inverted once, here, so a green cell always means strength. These are positions within a ranking, not accounting quantities: there is no cash-flow yield or margin figure on this page, because those are inputs the model does not publish.
Sectors
Utilities is genuinely homogeneous — 84% of it sits in a single profile. Industrials spreads across every one, its largest group holding only 22%. For a sector like that, the label says where a company is listed in a taxonomy, not how its business works.
Method
Every company in the release is placed by its 14 published readings — the nine analytical families, the four style scores and its size. Those are converted to van der Waerden rank-normal scores, reduced to 4 components by horn parallel analysis, 20 permutations (74% of the variance), and partitioned by k-means. The same release in always produces the same groups out: every random draw runs from a fixed seed.
This follows a research study that clustered the same universe on the model’s raw inputs. Those inputs are not published — which ones they are, and in what direction they act, is the part of the model that stays closed — so this works one level up, on the families the engine does publish. The structural findings hold: sector explains a small share of the variance, sub-industry more, the clusters more again, and the same sectors turn out homogeneous. The per-cluster cash-flow and margin figures of the original do not appear here, and no substitute for them has been invented.
Resampling 70% of the universe 20 times and refitting, the partition into 6 groups reproduces itself at an adjusted Rand index of 0.67. A coarser split into 3 is steadier still. 6 is used because it is the resolution the original study established and its groups do correspond — but the steadier figure belongs on the page next to it, not in a footnote. Read the six as a map of similarity, not as a classification with hard edges.
6 of the 949 companies in the release sit in Real Estate. Too few to form a group of their own, they would be either stranded as a cluster of their own or absorbed into one they have nothing in common with, so they are left out entirely — the same exclusion the original study made. 943 companies are grouped.
The clusters account for 57.5% of the variance against 12.0% for the 9 GICS sectors, but the clusters were fitted to maximise exactly that quantity and the sectors were not — on its own the comparison proves nothing. The 59 sub-industries are the real test: a hand-built taxonomy with 10 times as many groups as there are clusters, reaching 27.0%. That gap is the finding.
Each group is named from its own two strongest analytical families. Nobody writes these labels, which is why they read plainly — and why they stay correct after a release moves the boundaries. Group numbers run largest-first and are stable within a release, but not across releases: C3 next month need not be C3 today.
This page shows the structure. Inside ALB Equity, every one of the 943 companies carries its group, its distance from the centre of it, and the strongest names within each profile by the engine’s own ranking.