Quantifying Environmental Risks Using a Fusion of Statistical and Machine Learning Methods

Abstract

There is growing scientific evidence that adverse weather events are increasing in frequency and intensity. Such natural hazards can be largely separated into two groups. The first group comprises extreme events and natural disasters such as hurricanes, severe storms and tornadoes that due to their severity can be regarded as catastrophes. This group of extreme events is widely explored in statistical, climate, and engineering literature, and in practice is typically handled by the reinsurance industry. The second group of adverse atmospheric events includes the so-called low individual but high cumulative impact (LIHC) events, such as heavier than normal rain or stronger than usual wind speed. While these adverse events are not classified as catastrophes and their individual impact might be noticeable but low (e.g., roofs of a few houses in the neighborhood are damaged), the joint cumulative impact of such events is substantial and tends to increase over the years. Furthermore, LIHC events are largely unexplored from statistical, actuarial and engineering perspectives, and our society, with often outdated city critical infrastructures, remains vulnerable and unprepared to such climate scenarios of increasing LIHC events. In this thesis, first, we have investigated statistical and data science methods that can be used not only to model but also to predict the impact of LIHC events on the residential sector. We have developed a nonparametric LIHC risk analysis that allows to account for the impact of multiple exogenous atmospheric variables on the dynamics of home insurance claims. Second, we have explored the impact of climate dynamics on agriculture. Indeed, agricultural production systems are highly vulnerable to shorter-term extreme weather events and longerterm climate variability and change. Resulting environmental impacts drive broader economic and social impacts on farmers, insurers, and other stakeholders across agricultural supplychains. Improved multivariate statistical methods are urgently needed for quantifying crop yield risk driven by variables that are strongly spatially and temporally dependent, while also exhibiting more frequent extreme behavior. Third, we have investigated dynamics of (re)emerging climate sensitive infectious diseases. Nowadays with an ever-increasing globalization and world connectivity, it is broadly recognized that mitigation of risk of (re)emerging infectious diseases constitutes a global challenge rather than a problem of a single country or even a continent. Similarly, surveillance and forecasting of infectious diseases fall well beyond boundaries of a single discipline and require development of systematic multidisciplinary approaches at the interface of statistics, machine learning, epidemiology, and public policy. We attempt to fill some of the existing knowledge gaps in disease surveillance and forecasting, with a particular focus on influenza and Zika viruses, and explore how modern data science methodology can assist society in developing more accurate biosurveillance and prediction platforms. This dissertation aims to open up new horizons for studying the effect of climate changes and its impact on several human aspects, we incorporated statistical techniques that are not frequently used in forecasting the insurance claims and infectious diseases, showing that these techniques have promising results. We hope that the findings of this dissertation will help those sectors in their decision making process.