Custom industry classification of US companies
July 2025 – presentOfficial industry codes often say too little about what a company actually does. So I collect the annual reports of US listed companies myself, extract the business description and risk factors, and compare the texts. The result is a set of groups and themes that cut across industries, which I use to find related companies when a stock rises or falls unusually sharply.
An end-to-end example: Red Cat. Red Cat Holdings makes drones for the US military. Here we follow the company through the method.
What is Red Cat according to the official codes?
The SIC code puts Red Cat with software companies, and GICS puts it with aircraft manufacturers and defence giants. Neither tells you that it is a small drone company.
Only what is specific to Red Cat counts
The risk factor section of Red Cat's annual report is split into passages. Boilerplate – passages that appear almost word for word at many other companies – is removed, because otherwise it makes every company look alike. Most of Red Cat's passages turn out to be boilerplate.
The text becomes vectors that can be compared
passages→ embedding model→one vector per passage
and per document
An embedding model reads the text and turns it into a long vector of numbers, where texts with similar meaning get vectors that point in the same direction. This is done for each passage of the risk factor section, for the section as a whole and for the company profile. Alongside this, TF-IDF measures which words are distinctive for the company, and that word vector is compressed with SVD. Each vector is normalised, weighted and combined into a single vector per company.
Fingerprint: the first 24 principal components of each company's combined vector. Blue is below average, red above, grey close to zero. Red Cat and Ondas clearly share the same pattern; Microsoft's is different. Hover over a cell to see the value.
Cosine similarity between Red Cat and four other companies, for each source and combined. The highest value in each column is in bold.
The sources disagree. TF-IDF counts words and finds Kratos closest, because they share the language of defence contracts. Embeddings place companies by what the text means, and find Ondas closest. The company profile is the only text the company has not written itself in its annual report, and it points the same way. That is why the sources are weighted together rather than relying on any one of them.
Red Cat's nearest peers, ranked
Every company is compared with every other in three ways: the most similar text passages, TF-IDF and the business description. Each measure is standardised, and their sum gives a single score by which all companies are ranked. The table shows Red Cat's eight nearest peers, with their rank on each measure.
The two closest are the defence companies Kratos and AeroVironment, which share the language of Pentagon contracts. But Red Cat shares its drone vocabulary with AIRO, Ondas and Unusual Machines. So the individual peers pull in two directions. It is the groups in the next step that bring the picture together.
Every company, placed by how similar their text is
Hover over a point to see the company and its group. The map is a t-SNE projection of the same text vectors the groups are built from. The projection preserves which companies are close to each other, but not the exact distances between groups.
The groups are formed with two-level Ward clustering on the combined text vectors, and on those alone. Clusters are built from companies above a certain size, so that the many very small companies do not dictate the groups. Smaller companies are then assigned to the group they are closest to. Red Cat is large enough to be included from the start. A language model names the groups afterwards, but it plays no part in forming them.
The software companies from the SIC code sit somewhere else entirely. Red Cat, on the other hand, sits right among the GICS defence companies because the text is similar, and Red Cat's own group appears as two clusters. The map is only a two-dimensional sketch. The grouping happens in the full vectors, where Red Cat is closest to the drone group, as the next step shows.
Does Red Cat's own group fit better than SIC and GICS?
Returns play no part in forming the groups. They are used afterwards to check whether the grouping is right. Here Red Cat is compared with its own group and with the two official groups from step 1.
Each dot is one group. Text: how close Red Cat's text is to the group average. Returns: how closely Red Cat's daily returns track the group's average return over the past year. Hover over a dot to see the value.
Red Cat's own group is the closest on both rows. The SIC code misses by a wide margin: the software companies' returns track Red Cat poorly. GICS comes closer, because the defence industry does share customers with Red Cat.
How much do the groups move with the market?
Beta against the average stock over the past year, for each group as a whole and for Red Cat itself.
This is where the difference is clearest. Red Cat swings far more than the average stock, and only its own group comes close. Software and defence stocks swing much less. An investor comparing Red Cat with the official groups would get the wrong picture of the risk.
Where Red Cat started, and where it ended up
Before · SIC 7372
Software
- Microsoft software and cloud
- Oracle databases and cloud
- Palantir data analytics for government and business
- CrowdStrike cybersecurity
- SAP business software
Before · GICS
Aerospace & Defense
- GE Aerospace jet engines
- RTX jet engines, missiles and defence systems
- Boeing commercial and military aircraft
- Lockheed Martin fighter jets, missiles and space
- General Dynamics submarines, tanks and business jets
After · Red Cat's own group
Drone & Robotics Ventures
- Red Cat drones for military and commercial use
- Ondas drones and private wireless networks
- Unusual Machines small drones and drone components
- AIRO Group drones, avionics and pilot training
- Serve Robotics autonomous food delivery robots
- Richtech Robotics service robots for restaurants and more
- Palladyne AI AI software for robots
Red Cat started out alongside the world's largest software and defence companies. It now sits among small companies that make drones and robots and, like Red Cat, swing far more than the market.
From the wrong industry code to the right peers
The official codes grouped Red Cat with software companies and defence giants. By reading what the company writes about itself, removing boilerplate and comparing the text in three ways, Red Cat ended up in a group of small drone and robotics companies. Returns played no part in the grouping, but they confirm it: Red Cat tracks its own group more closely than either official one, and only that group comes close to how much Red Cat swings.
That Red Cat tracks its own group more closely is the whole point. Comparable companies are used everywhere in finance: to build portfolios, to value stocks and to explain where risk comes from. Get the peer group wrong, and all three go wrong. Researchers at BlackRock reached a similar conclusion: companies that are similar in embeddings of their business descriptions in SEC filings also have similar returns (Company Similarity using Large Language Models, 2023).
