Files
Abstract
This dissertation consists of three essays on corporate finance and machine learning. The first essay investigates the relation between CEO conscientiousness and reserve management in U.S. property-liability insurers. The psychology literature claims that conscientiousness is one of the strongest predictors of work-related behavior. I find that CEO conscientiousness is negatively associated with reserve errors in the upper tail of the conditional distribution (at 75th percentile and higher), indicating insurers with more conscientious CEOs reserve less than insurers with less conscientious CEOs at a higher level of reserve errors to lower the cost of excess reserve rather than conservatism when reserve errors are extremely conservative. The evidence also shows that the negative relation is mitigated when insurers face high financial risk. Furthermore, more conscientious CEOs reserve less than less conscientious CEOs after SOX (compared with before SOX) when insurers face higher financial risk, possibly because they are more responsible for financial statements. The evidence is consistent with one feature of conscientiousness: following the rules and norms. Finally, more conscientious CEOs are better rewarded than less conscientious CEOs.The second essay studies the relation between corporate opacity and net premium written as a proxy of policyholders’ purchase behavior in U.S. property-liability insurers. I find that policyholders are willing to buy policies from less opaque insurers. The evidence also shows that policyholders are more sensitive to information about insurers’ financial risk when they are less opaque. Additionally, policyholders are aware of insufficient protection by the guaranty fund. It further suggests that opacity significantly influences the purchase behavior of commercial lines, due to the involvement of brokers and agents who possess in-depth knowledge of insurers’ financial situations and product policies. Thus, insurers’ opacity plays a crucial role in shaping policyholders’ purchase behavior.The third essay applies machine learning methods to detect physicians. Physician fraud takes an important portion of healthcare fraud which needs continuous assessment and revision of the control methods. Using a large dataset from a life insurer in Taiwan, I construct 32 features and use multiple methods, including the neural network and RUSBoost methods to detect fraudulent physicians. Based on the neural network model, I further analyze the importance of features in detecting fraudulent physicians. Addressing the imbalanced data issue, the AUROC score of the neural network model is 0.781 for physicians with multiple claims. I find the cost savings range from 16.3% to 36.9% assuming the fraud rate of fraudulent physicians’ total claim amounts ranging from 30% to 70%. I also find the important features to identify fraudulent physicians are associated with physicians clustering in the eastern area of Taiwan, the percentage of insureds whose age are less than 18, the percentage of surgeries due to illness as opposed to accidents, and whether the physician can perform difficult surgeries. Finally, the evidence implies fraudulent physicians use the "steal a little, all the time" strategy to avoid being caught. Besides cost savings, this study can benefit the life insurer by speeding up the claim review process, narrowing down the investigation range, and excluding suspicious physicians as external reviewers.