Application of an interpretable deep neural network to econometric modelling in management science

Abstract

Research in the field of deep neural networks has made rapid progress in recent years and proved its potential. Deep neural networks models are excellent in capturing complex non-linear relationships between explanatory variables and dependent variable, but it cannot provide logical process or grounds of the predicted results. Among various bypass methodologies proposed to overcome this limitation, attention quantifies the importance of input features and provides intuitive interpretable clues. In this dissertation, the author not only quantifies the importance of input features and their uncertainty, but also explores the possibility of applying interpretable deep neural networks into business/economic quantitative analysis. This dissertation consists of five chapters and three main chapters. In the first main chapter, previous studies quantifying the importance of explanatory variables within Bayesian framework will be reviewed and their limitation will be discussed. The author proposes a new perspective that analogize the statistical significance of classical econometrics into deep learning context. In the second main chapter, by applying the proposed perspective, deep neural networks structure that can measure the impact of health checks on medical expenditure will be presented. The structure can detect and correct the self-selection bias and the author finds that ignoring self-selection causes substantial upward bias. It is confirmed that the health checks do not significantly affect the medical expenditure when self-selection bias is corrected. In the last main chapter, the author presents the deep neural networks models that can predict the inventory of online pre-roll video advertising or solve the profit maximization problem under asymmetric loss condition. The author finds that deep neural networks model concentrates on explanatory variable of low uncertainty and concentrates on the information of latest days adjacent to the target day. Based on these empirical analyses, it was proved that the deep neural networks models proposed in this dissertation can provide logical basis for decision-making as well as accurate forecasting.