Research

Working Papers

Model Complexity and Restrictiveness (with Keaton Ellis) 

presented at the 2025 California Econometrics Conference

We study the measure of restrictiveness proposed in Fudenberg et al. (2023) to evaluate complexity of economic models using synthetic data. We show that Rademacher complexity is an affine transformation of a particular case of a consistent finite-sample estimate of restrictiveness. We highlight that this connection comes with both desirable and undesirable traits regarding bounds on generalization error and (in)ability to distinguish between falsifiable models, which produces a tradeoff in measure selection dependent on the relative importance of each trait to the analyst.

Works in Progress

Model Completeness, Restrictiveness, and Excess Risk

I study the measures of model performance called completeness and restrictiveness from Fudenberg et al. (2023). Together these measures define a Pareto frontier, where one model is preferred over another if it is both more complete and more restrictive. Why are completeness and restrictiveness valuable, and what is a practical way to choose between two models when one is more complete and less restrictive? To answer these questions, I identify connections between these measures and the measures of approximation error and estimation error from computer science. These relationships provide two important insights. The first is why completeness and restrictiveness are important model qualities: if model A is more complete and more restrictive than model B, then model A will tend to have smaller excess risk. The second is that the method of Structural Risk Minimization (SRM) introduced by Vapnik can be used to trade off these qualities. Since Cumulative Prospect Theory (CPT) is more complete yet less restrictive than Expected Utility Theory (EUT), I demonstrate in an application how SRM can be used to select between them.

When and Why Do People Overfit?

Machines overfit when the model they learn is too closely tailored to the training data, capturing not only the underlying structure but also idiosyncratic fluctuations, resulting in the model not being able to generalize well to new data. I experimentally study whether people overfit when they learn models from data by testing whether people follow the same principles that machines use to prevent overfitting, such as early stopping, network reduction, expansion of the training data, and regularization.

(Sub)-Optimal Search

awarded Xlab Experimental Social Science grant

There are innumerable social and economic situations in which agents search for information. Despite its importance for economic outcomes, we know very little about how individuals actually trade-off the option to search with the option to choose a known alternative. To understand how agents actually search, I design an experiment in which participants complete dynamic search problems, and I use a random utility model to describe deviations from optimal search behavior. Initial pilot results suggest that there is significant heterogeneity in individual search behavior and that how close an agent is to optimizing is correlated with a measure of their cognitive ability.