Title: Actual events vs. perceived reporting: Modeling firm performance under environmental uncertainty using machine learning
Authors: Minh Nguyen - University of Hawaii at Manoa (United States) [presenting]
Abstract: Not all companies respond the same to natural disaster events. The aim is to investigate two ways that natural disasters affect firm performance: actual events vs. perceived reporting. We consider the billion-dollar weather and climate disasters in the United States as the actual events and the number of words related to natural disasters in the Management Discussion and Analysis section in Form 10-Ks filing by the U.S. public companies as the perceived reporting. The aim is also to compare the performances of classification and regression trees (CART) and neural networks with linear regression in predicting the performance of U.S. public companies under environmental uncertainty. The results show that both actual events and perceived reporting of natural disasters this year negatively affect the return on assets in the next year. Also, the actual natural disasters this year negatively affect the next year's sales growth. An important result of neural networks is that deeper networks do not ensure improving the predictive accuracy in predicting firm performance. Comparing CART, neural networks, and linear regression models, we find that CART and neural networks outperform linear regression models in predicting firm performance. This result is robust to any given firm performance criteria, split ratios, and prediction errors.