Many aspects of economic productivity, including gross domestic product measurements, have been steadily rising over the past three decades. Although one would assume that the products that are available in the economy have become more diverse, a recent study has suggested otherwise. The results may help economists rethink how they measure the health of the economy.

Measuring product diversity is challenging because it is difficult to collect the necessary information about every product in an economy each year. Although previous studies have tried to model the data, the methods were too simple to capture the whole picture. In the current study, the researchers used three different models to analyze product diversity in large US firms from 1997 to 2017.

The researchers used 10-K documents, which are comprehensive, annual reports that are filed by companies to describe their financial performance. The team compared the reports to a dictionary to understand the pattern of words used and utilized that information to model product diversity.

“You ask what words occur in the document and put them in a bucket. If two documents have the same word, they are more similar,” said Mark Bedau, a professor of philosophy in Reed College. “However, these measures are crude and our point is that some words should count more because they are rare in the corpus but common in the document. It then becomes a more sensitive measurement.”

In addition to counting the words and weighting them based on how rare they are, the study also used a neural network model that predicts words based on the context. For example, if the sentence “the ball is round” is present in the document, the neural network is given the input “the ball is” and is then asked to predict the word “round”. By doing so, the model ends up learning how similar the firm documents are, according to Ananthan Nambiar, a graduate student in the Maslov group.

The inspiration behind analyzing economic data using a biology-based measure came from an unlikely source—microbial communities. “I look at these communities in my work at the IGB,” Nambiar said. “One thing that comes up a lot is the community diversity. We decided to apply that principle here and use different ways of looking at diversity to see how it was changing.”

The study compared different firms that manufacture similar products, such as Johnson & Johnson, Merk, and Pfizer; AT&T and Verizon; Home Depot and Walmart; and IBM, Intel, Cisco and Microsoft. To everyone’s surprise, the diversity was falling. “It’s like going to the grocery store and seeing shelves filled with by products made by the same company and these products are more similar to each other. You basically have less choice as a consumer,” Bedau said.

Even though the decreased product diversity could be due to narrowing consumer demand or companies moving their business overseas, other reasons include the kind of data investigated in the study. “We only looked at publicly traded US firms and didn’t take small US firms into account,” Nambiar said. “There could also be products developed in large or small non-US firms that we didn’t look at.” Additionally, 10-K documents introduce an element of bias because they are shareholder reports. The researchers would, therefore, like to further corroborate these results with other product descriptions and more sophisticated models.

“This paper is a nice case study of how interdisciplinary work is important,” said Tobias Rubel, a graduate student in computer science at the University of Maryland. “Our domain is in economics and we’re using tools from data science and machine learning, as well as modeling techniques from computational biology.”

The study “Dropping diversity of products of large US firms: Models and measures” was published in PLOS ONE.