Assistant Professor of Economics and Computer Science Northwestern University (by courtesy)
310 Sutardja Dai Hall
September 16, 2022
Whether a model’s performance on a given domain can be extrapolated to other settings depends on whether it has learned generalizable structure. We formulate this as the problem of theory transfer, and provide a tractable way to measure a theory’s transferability. We derive confidence intervals for transferability that ensure coverage in finite samples, and apply our approach to evaluate the transferability of predictions of certainty equivalents across different subject pools. We find that models motivated by economic theory perform more reliably than black-box machine learning methods at this transfer prediction task.