Using machine learning to revisit the diversification–performance relationship

Abstract

Research Summary In this article, we examine the relationship between corporate diversification and firm performance using a machine learning technique called natural language processing (NLP). By applying a widely used NLP technique called topic modeling to unstructured text from annual reports, we create a new, multidimensional measure that captures the degree of diversification of both multisegment and single‐segment firms. Additionally, we introduce a novel method to incorporate human judgments into the interpretation of machine‐learned patterns, which allows us to measure diversification across multiple dimensions, such as products and geographies. Finally, we illustrate how these new measures can generate novel insights into the relationship between the degree and type of diversification and firm performance, furthering our understanding of the diversification–performance relationship. Managerial Summary At some point, most firms face dilemmas about whether to diversify their business activities across industries or geographic markets—an important decision that invariably affects firm performance. Albeit very important, the direction of a relationship between diversification and firm performance is not always clear. Inconsistent results of previous studies are partially driven by inherent difficulties in reliably measuring diversification. This study introduces a novel methodology to address that problem: a machine learning‐based technique to quantify diversification from unstructured corporate annual report texts. An analysis of firm performance based on these novel diversification measures suggests that diversification, in contrast to earlier studies that find a diversification discount, is associated with higher firm value—a premium particularly pronounced for firms diversifying within a single industry.

Framing the virus: The political, economic, biomedical and social understandings of the COVID-19 in Taiwan

“… 5 The commonly used rule in previous topic modeling studies is to filter out words that appear in less than 0.5 %–1 % of documents ( Denny and Spirling, 2018 ). Our choice of 1 % was consistent with Choi et al (2021) and Hickman et al (2022) , who argued that, with longer documents (compared with social media text), using higher thresholds can cut computational cost and hardly impact the result of the analysis. …”

“…We did so by applying the Jieba tokenizer, a Chinese word segmentation tool ( Chen and Hsu, 2018 ; Day and Lee, 2016 ). We then removed highly infrequent words that appeared in less than 1 % of documents 5 to reduce computational cost ( Choi et al, 2021 ; Hickman et al, 2022 ) and stop words (e.g., the, but, a, and, if) that contain little topical content.…”

“…More often, the question is to balance detail with abstraction ( Aranda et al, 2021 ). As DiMaggio et al (2013: 582) noted, “there is no statistical test for the optimal number of topics,” and that “the point is to identify the lens through which one can see the data most clearly.” In this sense, we variably adjust the number of topics to inspect the interpretability of the results — in particular, with a concern to see whether new topics emerge when we increase the number of topics ( Aranda et al, 2021 ; Choi et al, 2021 ; Haans, 2019 ; Jeong et al, 2019 ; Kaplan and Vakili, 2015 ). We utilize pyLDAvis package in Python, a visualization software program for topic modeling, to perform such qualitative assessments.…”

Using supervised machine learning for large‐scale classification in management research: The case for identifying artificial intelligence patents

“…Second, unsupervised ML approaches, such as topic modeling, have been used to identify dominant patterns or groups within text data or image data without the guidance of any preexisting theoretical concepts (Choudhury et al, 2019; Kaplan & Vakili, 2015; Choi, Menon, Tabakovic, 2021; Teodoridis et al, 2020; for a recent review, see Hannigan et al, 2019).…”

The resource‐based view, stakeholder capitalism, ESG, and sustainable competitive advantage: The firm's embeddedness into ecology, society, and governance

“…Social science has adapted machine learning, NLP, and CATA successfully, and a recent contribution in Strategic Management Journal (Choi et al, 2021) suggests that the modern methods are better predictors with accuracy and efficiency. While topic modeling is also gaining ground, CATA has been established since 2009 (Uotila et al, 2009) in strategic management discourse.…”