← Research

Asset similarity

Peers and diversifiers, when the sector label isn't the answer

Published 2026-06-26 · 2,068 names · 252d window · 25 clusters

Two stocks in the same sector can behave nothing alike, and two in different sectors can move as one. This study builds a similarity measure that ignores the industry label and blends two things instead: how a name behaves (market-residual returns, PCA-denoised) and what the company is (fundamental ratios and sector structure). Blended 0.55 / 0.45 across 2,068 common stocks, it produces a full similarity matrix and a 25-way partition of the universe. The useful output is not the peer list but its inverse: the names that are genuinely independent of what you already hold.

2,068
Names in the final similarity matrix
22 / 252
Eigenvalues above the noise bound, kept as signal
25
Clusters, sized 44–129
34 MB
Size of the resulting matrix

Getting to a clean universe

Most of the attrition is funds, not filtering for quality. Leaving ETFs in would have poisoned the behavioural block, because a fund's returns are a blend of its holdings and would look similar to everything.

  1. 01 Liquid names requested 3,000 names
  2. 02 Common stocks after dropping funds 2,141 names

    859 dropped: 738 ETF, 39 ETV, 9 ETS, 6 ETN, 3 PFD

  3. 03 With a full ~1y return window 2,068 names

The two blocks

Behaviour alone overfits to a single window; characteristics alone just rediscover the sector label. Blending them is what makes a cross-sector twin findable.

Behavioural

0.55 of the blend

Market-residual returns over 252 days, denoised by keeping only the 22 eigenvalues that clear the Marchenko-Pastur bound of 14.9 — everything below it is indistinguishable from noise in a matrix this shape.

λ1172λ2104.9λ372.8λ455.7λ549.4

Characteristic

0.45 of the blend

35 dimensions — 10 numeric fundamentals plus 25 sector indicators. Coverage is the weak point: ratios do not resolve for every name, so this half of the blend is thinner for loss-making companies.

Return on assets
87%
Price to earnings
63%

Worked example — AAPL

AAPL sits in electronic computers, but nothing in its own industry makes the top of its peer list. What it actually trades like is large, stable, cash-generative consumer and payments names.

Most similar

MA
0.61 Services — business services
V
0.59 Services — business services
CHRW
0.56 Freight arrangement
TJX
0.54 Retail — family clothing
GWW
0.54 Wholesale — durable goods
EXPD
0.52 Freight arrangement
CL
0.52 Cosmetics & toiletries
LLY
0.49 Pharmaceutical preparations

Most independent

LIDR
-0.54 Motor vehicle parts
RAL
-0.53 Industrial instruments
DFDV
-0.50 Finance services
MARA
-0.49 Finance services
SBET
-0.49 Finance services
PRCT
-0.48 Surgical & medical instruments
ENVX
-0.46 Electrical machinery
ABCL
-0.45 Pharmaceutical preparations

Structure of the universe

Clustering the matrix recovers groupings that mostly look sensible and occasionally do not — cluster 12 pulls together international large-caps and materials names that share little except a dollar sensitivity. Showing the six largest of 25.

Cluster 18 129 names

Consumer & travel cyclicals

NKEDISLULUCCLUALTGTDALCRH
Cluster 19 113 names

Mega-cap platforms

MSFTAMZNMETAORCLNFLXAPPVUBER
Cluster 14 109 names

Enterprise software

PLTRCRMNOWCRWDPANWINTUSNOWSHOP
Cluster 12 106 names

International & materials

NUVALEARGXRIOBTIONONRACENVS
Cluster 23 106 names

Biotech & pharma

INSMMRNARVMDLCIDBBIOMDGLIONSJAZZ
Cluster 2 104 names

Speculative frontier tech

CRWVRKLBOKLORGTIIONQASTSQBTSOPEN

Limitations

  • Similarity is measured over one 252-day window. Relationships that only appear in a different regime are invisible here, and the whole matrix shifts as the window rolls.
  • Characteristic coverage is uneven — price-to-earnings resolves for only 63% of names, so the characteristic half of the blend is thinner for loss-making companies.
  • Sector labels come from SIC codes, which are coarse and occasionally missing entirely. Several genuine peers show a blank sector.
  • The blend weights (0.55 behavioural / 0.45 characteristic) are a judgement call, not an optimisation. A different split reorders the middle of every peer list.
  • Clusters are a partition of this universe at this moment, not stable industry definitions. Names near a boundary move between runs.

Window 2025-07-28 to 2026-06-26. Blend weights 0.55 behavioural / 0.45 characteristic. Matrix 2068×2068 (34 MB).