Machine learning applications in agricultural economics

The European Union has established ambitious environmental targets for agriculture, including achieving 25\% organic farmland by 2030. However, evaluating whether current policies effectively promote sustainable farming practices remains challenging. Traditional econometric methods often struggle with the complexity of agricultural systems, where economic, environmental, and social factors interact in intricate ways.

This dissertation investigates how machine learning algorithms can improve agricultural policy evaluation. Machine learning refers to computational methods that can identify complex patterns in data without requiring explicit programming of all relationships. The research addresses a fundamental question: can these advanced analytical tools provide better insights into the effectiveness of sustainability policies in European agriculture?

The study employs multiple methodological approaches. First, it compares machine learning algorithms with traditional econometric methods for forecasting agricultural commodity prices, demonstrating superior performance in handling market disruptions. Second, through simulation studies using European farm data, it evaluates how different methods perform under various analytical challenges such as non-linear relationships and selection bias. Third, it develops a novel framework called Double/debiased Machine Learning Staggered Difference-in-Differences (DML-SDiD) that combines machine learning's flexibility with econometrics' causal inference capabilities.

Applying these methods to evaluate Common Agricultural Policy initiatives (2014-2020) yields important findings. Analysis of agri-environmental programmes reveals widespread windfall effects, where participating farmers often maintain existing practices despite receiving payments. The evaluation of organic farming conversion uncovers more complex dynamics: farmers strategically reduce input expenditures during the transition period when organic price premiums are unavailable, then adjust management practices after certification. These patterns vary significantly across farming systems, with livestock operations showing minimal disruption while crop farms undergo substantial adaptations.

The research demonstrates that machine learning can enhance policy evaluation in three ways: by handling complex non-linear relationships between variables, by maintaining statistical validity even with limited sample sizes, and by revealing heterogeneous policy effects across different farming contexts. However, these methods require careful implementation guided by domain expertise and theoretical understanding.

This work contributes to both methodological advancement and policy insight. Methodologically, it provides frameworks and best practices for implementing machine learning in agricultural economics. Practically, it offers evidence that current sustainability policies may not achieve intended behavioral changes, suggesting the need for refined policy design. The findings indicate that while potential environmental benefits from organic conversion might be substantial, the economic challenges faced by converting farms underscore the importance of continued public support.

By bridging advanced computational methods with agricultural policy analysis, this dissertation enables more nuanced understanding of how farms respond to sustainability initiatives, ultimately supporting evidence-based policymaking for European agriculture's environmental transition.