Economists speed up a core business tool by replacing its restrictive assumptions with neural networks

A new method lets economists build choice models that don't rely on oversimplified assumptions about how people make decisions. Instead of fitting data to a formula that only works if you pretend errors are uncorrelated, you train a neural network to approximate the real probabilities, then use that to do the statistical work. This matters because it means better predictions of consumer behavior without the math getting intractable.

For decades, applied economists and marketers have used logit models because they were the only option that didn't require you to run expensive simulations just to evaluate a likelihood function. That meant accepting false assumptions about substitution patterns and error correlation. This approach removes that tradeoff: you get realistic error structures and you get fast computation. The practical effect is that choice models become less of a compression of reality and more of an actual approximation of it.