Risk Analytics in Econometrics

Abstract

[eng] This thesis addresses the framework of risk analytics as a compendium of four main pillars: (i) big data, (ii) intensive programming, (iii) advanced analytics and machine learning, and (iv) risk analysis. Under the latter mainstay, this PhD dissertation reviews potential hazards known as “extreme events” that could negatively impact the wellbeing of people, profitability of firms, or the economic stability of a country, but which also have been underestimated or incorrectly treated by traditional modelling techniques. The objective of this thesis is to develop econometric and machine learning algorithms that can improve the predictive capacity of those extreme events and improve the comprehension of the phenomena contrary to some modern advanced methods which are black boxes in terms of interpretation. This thesis presents seven chapters that provide a methodological contribution to the existing literature by building techniques that transform the new valuable insights of big data into more accurate predictions that support decisions under risk, and increase robustness for more reliable and real results. This PhD thesis focuses uniquely on extremal events which are trigged into a binary variable, mostly known as class-imbalanced data and rare events in binary response, in other words, whose classes that are not equally distributed. The scope of research tackle real cases studies in the field of risk and insurance, where it is highly important to specify a level of claims of an event in order to foresee its impact and to provide a personalized treatment. After Chapter 1 corresponding to the introduction, Chapter 2 proposes a weighting mechanism to incorporated in the weighted likelihood estimation of a generalized linear model to improve the predictive performance of the highest and lowest deciles of prediction. Chapter 3 proposes two different weighting procedures for a logistic regression model with complex survey data or specific sampling designed data. Its objective is to control the randomness of data and provide more sensitivity to the estimated model. Chapter 4 proposes a rigorous review of trials with modern and classical predictive methods to uncover and discuss the efficiency of certain methods over others, and which and how gaps in machine learning literature can be addressed efficiently. Chapter 5 proposes a novel boosting-based method that overcomes certain existing methods in terms of predictive accuracy and also, recovers some interpretation of the model with imbalanced data. Chapter 6 develops another boosting-based algorithm which is able to improve the predictive capacity of rare events and get approximated as a generalized linear model in terms of interpretation. And finally, Chapter 7 includes the conclusions and final remarks. The present thesis highlights the importance of developing alternative modelling algorithms that reduces uncertainty, especially when there are potential limitations that impede to know all the previous factors that influence on the presence of a rare event or imbalanced-data phenomenon. This thesis merges two important approaches in modelling predictive literature as they are: “econometrics” and “machine learning”. All in all, this thesis contributes to enhance the methodology of how empirical analysis in many experimental and non-experimental sciences have being doing so far.