From Predictive Algorithms to Automatic Generation of Anomalies

Sendhil Mullainathan, Ashesh Rambachan

TL;DR

This paper introduces a systematic way to use machine learning to improve economic theories by automatically generating "anomalies"—specific scenarios where a theory’s predictions fail. By pitting a restrictive economic theory against a high-performing "black-box" predictive model in an adversarial framework, the authors successfully recovered famous historical paradoxes and discovered entirely new categories of human decision-making failures that were subsequently validated in real-world experiments.

What Problem Does This Paper Address?

In economics, theories are often improved when researchers identify "anomalies"—counter-examples like the Allais Paradox that prove a reigning theory (such as Expected Utility Theory, or EUT) is fundamentally flawed. Historically, finding these anomalies required the creative intuition of brilliant researchers.

Meanwhile, modern machine learning (ML) models can predict human behavior with much higher accuracy than traditional economic theories, but they are "black boxes" that don't offer theoretical insights. The authors bridge this gap by asking: How can we use the predictive power of ML to automatically find the flaws in our theories? They transform the "creative act" of anomaly generation into a formal optimization problem.

Methods

The authors treat a theory as an allowable function class—a set of rules that restrict what outcomes are possible. They then propose two main procedures to find where these rules break:

  1. Adversarial Procedure: This is framed as a max-min game. A "falsifier" (the algorithm) searches for a collection of examples that maximizes the "loss" of the theory. Essentially, it hunts for data points where even the best possible version of the theory cannot match the predictions of a high-accuracy ML model.
  2. Representational Morphing: Many theories assume certain different inputs are "representationally equivalent" (e.g., EUT’s independence axiom implies that adding the same "side bet" to two different lotteries shouldn't change your preference). The algorithm uses gradient descent to find "nearby" scenarios that the theory thinks are identical but the ML model predicts will cause a choice reversal.

Datasets

The authors validated their approach using three distinct data sources:

  • Simulated Data: They first used data generated by Prospect Theory (a known alternative to EUT) to see if their algorithm could "re-discover" classic anomalies like the Allais Paradox.
  • Choices13K: A large-scale public dataset (Peterson et al., 2021) containing over 1 million decisions across 13,006 lottery menus. They used this to train a shallow neural network (2 hidden layers, 32 nodes each) to serve as their "empirical intuition."
  • New Experimental Data: To prove the machine-generated anomalies weren't just artifacts of the model, the authors recruited ~1,000 participants on Prolific for incentivized experiments, testing 60 of the newly discovered anomalies with real money at stake.

Key Insights and Ideas

  • Anomalies as Adversarial Examples: The core "aha moment" is realizing that an economic anomaly is functionally identical to an "adversarial attack" in computer vision. Just as a few pixels can trick a neural net into seeing a pig as an airplane, a specific lottery menu can "trick" a theory into making a false prediction.
  • ML as a Stepping Stone: Instead of viewing ML as a replacement for theory, the authors view it as a source of intuition. The ML model points to where the "weirdness" in the data is, and the anomaly generation procedure crystallizes that weirdness into a testable hypothesis.
  • Automated Discovery of the "Blind Spots": The algorithm found that humans violate EUT in ways researchers hadn't formally categorized before, such as the "Strict Dominance Effect" and the "Dominated Consequence Effect."

Key Results

  • Recovery of Classics: The algorithms successfully re-discovered the Allais Paradox and the Common Ratio Effect without being told to look for them.
  • Novel Anomalies: The procedures generated hundreds of new anomalies. For 2-payoff lotteries, they found 881 verified anomalies; for 3-payoff lotteries, they found 1,996.
  • Empirical Validation: In new experiments, humans violated EUT on these algorithmically generated menus at rates of 7.2% to 14.5%. These violation rates are statistically significant and comparable in magnitude to the most famous paradoxes in behavioral economics (which typically see ~13-16% violation rates in similar incentivized settings).

Limitations and Future Work

  • Optimization Challenges: The search for anomalies is a non-convex optimization problem. While the authors used Gradient Descent Ascent (GDA), they acknowledge that more sophisticated optimization routines might find even "stronger" anomalies.
  • Interpretability: While the machine finds the anomaly, a human researcher is still needed to look at the results, categorize them, and give them names (like "Strict Dominance Effect"). Future work could involve using LLMs or other techniques to automate the interpretation of the generated anomalies.
  • Scope: The method requires "ample data" to train a predictive model and a theory that makes "sharp predictions." It remains to be seen how well this scales to more complex, multi-dimensional economic domains like game theory or macroeconomics.

Summary updated

Abstract

How can we extract theoretical insights from machine learning algorithms? We take a familiar lesson: researchers often turn their intuitions into theoretical insights by constructing "anomalies" -- specific examples highlighting hypothesized flaws in a theory, such as the Allais paradox and the Kahneman-Tversky choice experiments for expected utility. We develop procedures that replace researchers' intuitions with predictive algorithms: given a predictive algorithm and a theory, our procedures automatically generate anomalies for that theory. We illustrate our procedures with a concrete application: generating anomalies for expected utility theory. Based on a neural network that accurately predicts lottery choices, our procedures recover known anomalies for expected utility theory and discover new ones absent from existing work. In incentivized experiments, subjects violate expected utility theory on these algorithmically generated anomalies at rates similar to the Allais paradox and common ratio effect.
BibTeX
@misc{mullainathan2024from,
  author = {Sendhil Mullainathan and Ashesh Rambachan},
  title = {{From Predictive Algorithms to Automatic Generation of Anomalies}},
  year = {2024},
  eprint = {2404.10111},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.10111},
}

Loading related content...

Figure