Abstract

This dissertation is composed of three manuscripts (manuscript 1 has been published in Energy Economics, manuscript 2 is currently under review with the Journal of Energy Markets, and manuscript 3 is currently under review with the Obesity Research and Clinical Practice journal) that address problems in energy economics, finance, and epidemiology through statistical modeling.

The first manuscript proposes a covariate-dependent mixture model to describe the behavior of electricity DART spreads (defined as the difference between the day-ahead and real-time prices of electricity). The model incorporates multiple regimes and allows covariates to impact both the frequency and severity of DART spread spikes. Using data from the Long Island zone of the New York Independent System Operator, the model demonstrates a strong fit. Results reveal that including covariates in the severity component of the model is crucial, while mild additional performance is obtained with their inclusion in the frequency component. Neural network-based quantile regression benchmarks are unable to improve performance over our mixture model.

The second manuscript examines the diversification benefits of energy commodities during turbulent periods such as those marked by the COVID-19 pandemic and the Russia-Ukraine war, both of which deeply affected energy markets. Revisiting classical allocation strategies, we incorporate electricity futures—a rarely used asset—alongside crude oil and natural gas futures. Using mean-variance optimization, the diversification benefits are evaluated by combining these energy contracts with the S&P 500. Our empirical approach handles the non-stationarity of returns, volatilities, and correlations. Out-of-sample results show improved performance and diversification, especially during crisis periods.

The third manuscript extends existing dual-energy X-ray absorptiometry-based body composition classifications by introducing additional centile cut-offs to capture tail behavior. Using NHANES (National Health and Nutrition Examination Survey) data, we study the association between these phenotypes and health risks, including metabolic syndrome (MetS), depression, sleep disorders, and comorbidities. Nine phenotypes were identified using quantile regression (QR), and logistic regression was used to assess their relationship with health risks, compared to standard adiposity measures like body mass index (BMI), waist circumference (WC), and total fat percent. The QR model has a better (higher) LR+ (positive likelihood ratio) than the median-split model for MetS and comorbidity but consistently underperforms in LR- (negative likelihood ratio) compared to the median-split model. Both models perform worse than BMI and WC. Whether results differ over time or among certain subpopulations should be investigated.