The Impact of Machine Learning on Economics

Innovation and Economic Growth and Artificial Intelligence

Article Snapshot


Susan Athey


in The Economics of Artificial Intelligence: An Agenda, Ajay K. Agrawal, Joshua Gans, and Avi Goldfarb, eds., University of Chicago Press, 2019


Machine learning (ML) has begun to transform economics. ML systems can make it easy to classify large amounts of data. Empirical economists can use ML to develop new methodologies, ensure that others can reproduce their studies, and design more sophisticated studies.

Policy Relevance

Machine learning will have a dramatic effect on economics.

Main Points

  • ML develops algorithms designed to be applied to datasets, focusing on developing tools for making predictions and for classifying or grouping data.
  • “Unsupervised” ML identifies clusters of similar observations, and is used for video, images, and text.
    • If one reads that a researcher "discovered cats on YouTube," he might have used an unsupervised ML to divide a set of videos into groups, then watched the videos, discovering that most of the largest group contain cats.
    • This type of ML is "unsupervised," in the sense that the input data was not labelled.
  • Unsupervised ML could be used to find similar newspaper articles, online product reviews, and other materials, and classify the materials to identify variables for research.
    • A study of factors that affect consumer demand could use unsupervised ML to classify consumer product reviews (input variables).
    • A study to determine the effects of shutdown of Google News on consumers’ choice of news articles could use unsupervised ML to find patterns of news consumption (output variables).
  • “Supervised” ML uses a set of features to accurately describe ("predict") a new dataset, starting with some labelled observations in a separate ("training data") set; such a system could be used to accurately determine whether a set of images contains a certain animal.
  • New research explores how to use ML to solve causal inference problems, which differ from prediction problems.
    • Hotels raise prices when they are full, so one hotel could use price data to estimate the occupancy rate of a competitor, a prediction problem.
    • Although high prices are correlated with high occupancy rates, no one would predict that raising room prices would cause occupancy rates to rise.
    • Predicting how an increase in price would affect occupancy requires a more sophisticated study to support causal inferences.
  • Studies designed to choose optimal medical treatments must balance the desire to avoid giving patients suboptimal treatments with the need to learn; ML-based methods can actively learn which treatment is best, using data from early arrivals to assign treatments to later arrivals, and yield answers more quickly and reliably than randomized controlled trials.
  • ML can help economists solve matrix completion problems, which arise when some entries are missing from a data set; ML-based methods work better than traditional methods.
  • ML will change economic research in many ways, including the following:
    • Allowing use of off-the-shelf ML for tasks such as classification of data.
    • Enabling the development of entirely new methods by economists working with data on a large scale.
    • Inspiring research on the impact of artificial intelligence and ML on the economy.
  • Many empirical economists will adopt ML methods, because use of ML will better enable other economists to reproduce their results, and because ML perform tasks faster and more reliably than research assistants.
  • Economists worry about whether they have mistaken correlation for causation, whether they have considered hidden variables that affect the variables being studied ("cofounders"), and other problems relating to credibility of their work; machine learning will automate some routine data analysis tasks, but economists must continue to structure studies carefully.


Get The Article

Find the full article online

Search for Full Article