A company’s official sector does not always reveal which businesses most closely resemble it financially. In a study of 500 S&P 500 companies, the best machine-learning model predicted sector membership from accounting ratios with 49.3% validation accuracy. That was far above a simple 14.8% benchmark, but it also revealed substantial overlap between sectors.
The finding comes from a study by researchers at three Spanish universities, published in The Journal of Finance and Data Science. The paper became available online in May 2026 and was highlighted by its publisher on August 27.
The researchers did not conclude that stock market sectors are obsolete. Instead, their results suggest that a sector and a financial peer group answer different questions.
Sector labels describe businesses, not balance sheets
Companies in the S&P 500 are divided into 11 sectors under the Global Industry Classification Standard, commonly known as GICS.
GICS classifies a company according to its principal business activity. Revenue is the main consideration, although earnings and market perception can also influence the decision.
That makes the system useful for identifying businesses exposed to similar products, customers, regulations and economic conditions. It was not designed to group companies according to debt, liquidity, margins or cash generation.
The new study tested what happens when companies are viewed through that second lens.
Ricardo Reier Forradellas of the Catholic University of Ávila conducted the research with David Sánchez Cabrera of Spain’s National University of Distance Education, Luis Miguel Garay Gallastegui of the International University of La Rioja, and Sergio Luis Náñez Alonso of the Catholic University of Ávila.
They used fiscal 2022 financial statements for 500 unique S&P 500 companies. The figures were obtained through Alpha Vantage, which structures information drawn from company filings with the US Securities and Exchange Commission.
The team calculated ratios covering profitability, leverage, liquidity, interest coverage, operating efficiency and cash generation. It then tested seven standard machine-learning models to see how accurately those ratios could recover each company’s official sector.
The models found a sector signal, but not a clean divide
The strongest model was K-nearest neighbors, or KNN. The method classifies an observation by looking at the most similar observations around it. In this case, it assigned companies to sectors by comparing their financial profiles with those of nearby companies in the data.
KNN achieved 50.3% accuracy on the training data and 49.3% on the validation data, which had been held back to test the model.
The 49.3% result should not be treated as equivalent to a coin toss. The model was choosing among 11 sectors, and a naive model that labeled every company as an Industrial would have been correct only 14.8% of the time. Other measures designed for unevenly sized categories also showed that the accounting ratios contained meaningful information.
However, the signal varied sharply by sector. Real estate, utilities and financials each recorded validation accuracy above 70%, while the model failed to identify communication services companies successfully in the validation sample.
That unevenness is economically plausible. Utilities usually own extensive physical infrastructure. Banks have balance sheets that work differently from those of manufacturers or retailers. Such characteristics leave recognizable marks in their accounts.
Across the index as a whole, though, the financial boundaries were much less distinct. A company’s sector explained part of its accounting structure, but not all of it.
The researchers then removed the sector labels and asked an unsupervised clustering system to group the companies according to financial similarity alone. The process produced nine clusters that crossed conventional sector boundaries and generally showed lower internal variation across most of the ratios examined.
One group included Morgan Stanley, T-Mobile US and Warner Bros. Discovery. Another included American Airlines, Deere & Company and Disney.
These combinations do not mean the companies are competitors or broadly interchangeable. They mean that, in the fiscal 2022 data and across the selected ratios, they shared particular combinations of leverage, capitalization, operating costs, liquidity and cash generation.
One company can have several useful peer groups
The study supports a practical conclusion: the most useful peer group depends on the question being asked.
A company’s sector and industry remain logical starting points when examining competition, pricing power, regulation or customer demand. A software business is more likely to compete with other software providers than with an airline that happens to have similar leverage.
But an analyst studying debt capacity, short-term liquidity or financial resilience may learn more from companies with comparable balance sheets and cash-flow profiles, even when those businesses operate in different industries.
Valuation work may require both perspectives. An operational peer can help establish how the market values a particular type of business, while a financial peer can show whether differences in debt, margins or cash generation justify part of the valuation gap.
This is an inference from the study rather than a result it directly tested. The researchers did not examine investment returns, forecast stock prices or demonstrate that their clusters produce better valuations.
“Our findings do not mean that sector classifications are obsolete. They show that sectors tell only part of the story,” Forradellas said in the publisher’s research announcement.
Financial peers can change more quickly than industries
The accounting-based groups were not permanent. When the researchers compared later annual reporting periods, the retention measure declined from 82% for 2022 to 67% for 2023 and 61% for both 2024 and 2025.
That movement is not surprising. A company’s principal business may remain stable while its finances change after an acquisition, a debt repayment, a restructuring or a shift in profitability.
It also means accounting-based peer groups would need regular updating. A company that resembles one set of businesses today may resemble a different set after its balance sheet or cash generation changes.
The findings should still be treated cautiously. The main analysis used one year of financial statements and one 70/30 training and validation split. It covered only large US companies, relied on a selected set of ratios and used consolidated accounts that can obscure differences inside diversified businesses.
The nine-cluster result was also an economically interpretable compromise rather than a uniquely correct new classification of corporate America.
Even with those limitations, the study makes a useful distinction. A sector label explains the market in which a company operates. Its financial statements reveal how the business is currently funded and performing.
For investors and analysts, the closest competitor and the closest financial peer may therefore be two different companies.
