Identifying inter-firm relationships is critical in understanding the industry landscape. However, due to the dynamic nature of such relationships, it is challenging to capture corporate social networks in a scalable and timely manner. To address this issue, this research develops a framework to build corporate social network representations by applying natural language processing (NLP) techniques on a corpus of 10-K filings, describing the reporting firms’ perceived relationships with other firms. Our framework uses named-entity recognition (NER) to locate the corporate names in the text, topic modeling to identify types of relationships included, and BERT to predict the type of relationship described in each sentence. To show the value of the network measures created by the proposed framework, we conduct two empirical analyses to see their impacts on firm performance. The first study shows that competition relationship and in-degree measurements on all relationship types have prediction power in estimating future earnings. The second study focuses on the difference between individual perspectives in an inter-firm social network. Such a difference is measured by the direction of mentions and is an indicator of a firm’s success in network governance. Receiving more mentions from other firms is a positive signal to network governance and it shows a significant positive correlation with firm performance next year.
In this paper, we build on the network structural hole concept of organizational theory to theorize an individual firm-specific strategic competitive positioning (SCP) construct. We use unsupervised document embedding approaches to operationalize the SCP construct by capturing each firm’s relative competitive and strategic positioning in a strategic similarity matrix of all existing U.S. publicly traded firms’ annual corporate filings. This approach enables us to construct a theoretically driven firm-level SCP measure with minimal human expert intervention. Our construct dynamically captures competitive positioning across different firms and years without using artificially bounded and often outdated industry classification systems. We illustrate how the dynamic measure captures industry-level and cross-industry strategic changes. Then, we demonstrate the effectiveness of our construct with an empirical analysis showing the imprinting and dynamic effects of SCP on firm performance. The results show that our dynamic SCP measure outperforms existing competition measures and successfully predicts post-IPO performance. This paper makes significant contributions to the information systems and organizations literatures by proposing an organizational theory-based unsupervised approach to dynamically conceptualize and measure firm-level strategic competitive positioning. The construct can be easily applied to firm-specific, industry-level, and cross-industry research questions in many contexts across many disciplines.
Data: litigation risk scores for 6134 firms in 1996-2015 [link]
Research assistant: Raymond Situ
This study examines whether and how machine learning techniques can improve the prediction of litigation risk relative to the traditional logistic regression model. Existing litigation literature has no consensus on a predictive model. Additionally, the evaluation of litigation model performance is ad hoc. We use five popular machine learning techniques to predict litigation risk and benchmark their performance against the logistic regression model in Kim and Skinner (2012). Our results show that machine learning techniques can significantly improve the predictability of litigation risk. We identify two best-performing methods (random forest and convolutional neural networks) and rank the importance of predictors. Additionally, we show that models using economically-motivated ratio variables perform better than models using raw variables. Overall, our results suggest that the joint consideration of economically-meaningful predictors and machine learning techniques maximize the improvement of predictive litigation models.