UC Berkeley

Climate change may influence financial market participants in many ways. Particularly, in a market with financial frictions, real estate usually serves explicitly or implicitly as a collateral in debt financing. Risks of physical damage to real property resulting from climate hazards and sea level rise may bring about not only direct loss, but also credit constraint for property holders. How much agents are affected by climate risks is an important research question, and has been explored substantially in the burgeoning climate risk literature. Another important question is what actions can be taken to manage or reduce the risks, and how to evaluate those efforts, which is the main goal of this dissertation.

The first chapter studies the effectiveness and efficiency of adaptation investments for averting property damage (e.g. defensible space, drainage system, shoreline stabilization, etc.). Using variation in grant availability for adaptation projects through a U.S. federal program as a quasi-experiment, I quantify the impacts of property-related adaptation investments on debt financing of local governments and the real estate sector. There are three main findings. First, following adaptation investments, the average borrowing cost of a county government decreases by 10-26 basis points for 20 years. Second, nationally, adaptation investments have an insignificant effect on outstanding debt, while in the South and Northeast debt falls by 4.2%. Third, an average investment has a project cost of $2 million and reduces property damage by $323,000 per year, which implies a 15-year internal rate of return of 19%. Overall, these results suggest that adaptation mitigates climate risks. Additional calculations reveal that current levels of adaptation are below the social optimum; and given current spending, capital could be allocated more efficiently by altering the distribution across regions.

The second chapter leverages tools in natural language processing (NLP) to explore the potential of generating large-scale yet granular measurements of how individuals perceive climate change and actions for addressing climate challenges. Social media such as Twitter provide a platform for users with diverse backgrounds to freely share their opinions, and thus capture real-time, higher-dimensional information that is not reflected in standard opinion surveys or polls. In this essay, I evaluate the use of different machine learning models to classify opinions on climate change and related actions from tweets. For model training, I annotate a dataset of climate-related tweets using a multi-stage system that distinguishes between two types of climate actions, mitigation or adaptation. I show that a deep learning approach based on contextual embeddings (BERT) outperforms traditional models, and addressing unbalanced classes through up-sampling achieves additional gains in accuracy. Finally, I discuss the limitations and potential applications of text-based characterization of opinions on climate change actions.