This dissertation develops and applies computational methods to three problems at the frontier of empirical economics, united by a common objective: to operationalize abstract economic concepts using modern machine learning tools, to uncover new empirical knowledge through those measurements, and to evaluate the conditions under which such tools can be trusted.

The first essay asks what makes a scientific paper novel, and whether novelty of different kinds matters differently for impact. Treating novelty as a multidimensional construct rather than a single attribute, the essay leverages Large Language Models (LLMs) and embedding models to measure novelty directly from the full text of scientific papers. It finds that specific dimensions of novelty, defined by a paper's distinctiveness from its intellectual neighbors versus its engagement with the current frontier, predict fundamentally different outcomes. This finding reveals a strategic tension in knowledge production that citation-based measures of impact cannot capture. The second essay asks why strategic behavior systematically departs from game-theoretic predictions in complex environments. Drawing on a massive dataset of hundreds of millions of chess games, it develops an interpretable measure of when a position is genuinely difficult for a boundedly rational agent, and uncovers a new stylized fact: complexity arises endogenously more often among higher-skilled players. A structural model reveals that this pattern reflects experts' ability to sustain complexity, rather than a preference for risk, a finding with broader implications for understanding how people create and exploit strategic complexity. The third essay asks how much we should trust a predictive model when it is applied outside the context in which it was estimated. It formalizes this out-of-domain prediction problem, derives coverage-guaranteed forecast intervals for transfer error, and finds suggestive evidence that the in-domain superiority of black-box algorithms does not reliably generalize across domains, a cautionary note with direct relevance for the use of machine learning in situations requiring generalizability.

Together, these essays argue that the value of computational tools in economics lies not only in their raw predictive power, but more importantly in their capacity, when disciplined by economic theory, to render previously unobservable quantities measurable and to rigorously evaluate the limits of predictive models.