Predicting Household Welfare Outcomes Using Observable Socio-Economic Characteristics

Abstract

This thesis explores the predictive capacity of machine learning models to esti mate household welfare outcomes using socio-economic characteristics within the context of Index-Based Livestock Insurance (IBLI). IBLI aims to mitigate the economic shocks of climate-induced livestock loss by offering insurance based on satellite data. While previous studies have shown the heterogeneous welfare effects of IBLI, this thesis investigates whether machine learning methods are effective at uncovering heterogeneous welfare effects across household subgroups and can show the most important household characteristics. Four models have been evaluated: Lasso Regression, TabTransformers, Gen eralized Random Forest, and Bayesian Ridge Regression, using cattle and goat datasets. These datasets originate from a dataset with information about herders in southern Ethiopia who make IBLI purchase decisions for their herd. Results show that all models performed better on the goat dataset, with TabTransformers outperforming others in terms of predictive power. Despite modest overall per formance, key features such as the settlement status of households and trust in village insurance promoters consistently emerged as key predictors. These find ings highlight both the capabilities and limitations of applying machine learning in these contexts.