Economists build neural networks that predict how people choose faster than the old math does

Economists have a math problem: the standard models for predicting human choice (logit models) are fast but unrealistic. They don't capture how people actually substitute between options. This paper proposes using a trained neural network to approximate choice probabilities for messier, more realistic error patterns — and then proves the estimates are statistically valid and much faster to compute.

For decades, economists chose between two bad options: use simple math that runs fast but misses real behavior, or use complex simulations that get it right but take hours per estimate. This is a software solution, not a conceptual one, but it matters because it removes the speed penalty from realism. If this works in practice on real datasets, it means choice models in marketing, labor economics, and policy analysis can get more accurate without getting slower — which changes what questions researchers can afford to ask. The proof that the estimates remain statistically valid under imperfect approximation is the actual load-bearing part; it means you don't need the neural network to be perfect, just good enough.