Industries as Networks and Unspecified Lending Arrangements: Big Data and Machine Learning Applications

Abstract

In this thesis, we first develop a labor mobility network across industries (LMNInd) utilizing big data from a Chinese online job board and machine learning (ML) techniques, based on the concept of cross-industry skills (CRISs). The LMNInd is used to demonstrate the significance of CRISs in job search and reveal an absorbed ripple effect from industry-specific shocks. We then apply these ML techniques to compare labor mobility networks across industries versus occupations in the U.S. using Current Population Survey (CPS) data. This not only illustrates the robustness of the ML techniques but also showcases its broad applicability. Finally, we explore private lending markets through another ML approach, analyzing millions of legal documents related to private lending disputes with a large language model. This reveals a unique characteristic of private lending markets that around half of loans do not explicitly specify loan terms or rates.

In Chapter 1, we consider the labor market linkages across industries based on the concept of CRISs here. CRISs are skills that are productive beyond any single industry. CRISs connect industries through their impacts on workers' mobility hurdles across industries. In particular, we empirically estimate LMNInd using online job board data and recently available machine learning algorithm. Based on the estimated LMNInd, vacancy-applicant skill match indices are then constructed and tested on individual job application outcomes. We further demonstrate how this estimated LMNInd can be used to predict the transmission of an exogenous shock on one industry to all the other industries - the "ripple effect". Our results show a one standard deviation increase in the vacancy-applicant skill match index is associated with 0.48 percentage points increase in callback probabilities, which is equivalent to 1.5 (1.4) times of the impacts of being more educated (experienced) than required. The results also suggest that the effect of the vacancy-applicant skill match index is stronger for lower-paying jobs. Lastly, our aggregate level results suggest the so-called "ripple effect" can be non-linear and complicated due to the existence of CRISs and the specific framework of LMNInd in an economy.

In Chapter 2, we extend the methodologies from Chapter 1 to investigate labor mobility across industries and occupations in the US using CPS data. We compare labor mobility across industries versus that across occupations and find that they have similar patterns in mobility frequency. We further empirically estimate labor mobility network across industries (LMNInd) and occupations (LMNOcc). Based on the estimated LMNInd (LMNOcc), we forecast individuals' working industry (occupation). Our analysis of the predicted and realized industry (occupation) highlights the prediction power of our methodology, evidencing with over 90 percent accuracy. Lastly, we simulate industry (occupation) shocks using the estimated LMNInd (LMNOcc) to predict the effect on labor mobility across industries (occupations) and our results shed light on the ripple effects of industry (occupation) shocks.

In Chapter 3, we investigate the lending and borrowing behaviors in private lending markets. This informal financial market is an integral but often overlooked segment of the financial system. Our analysis is based on an innovative dataset, derived from legal sentencing documents in China and processed via a state-of-the-art large language model. By investigating the fundamental loan characteristics, such as loan size/term/rate, we find around half of loans do not explicitly specify loan terms or rates. We find heterogenous correlations between the unspecified lending arrangements and loan-related characteristics. We further investigate the correlation between the unspecified lending arrangements and the fundamental loan characteristics and find that larger loans are often associated with loans with less unspecified lending arrangements. Show more