ECONOMETRICS AND MACHINE LEARNING: THERE IS MORE TO IT THAN MEETS THE EYE

The very first question, an economist usually asks when encounters machine learning (ML) is what’s the difference between ML and econometrics. The right answer is probably – tradition. While both set of tools are based on the same set of statistical insights and very similar at the core, they were developed separately by different communities with different purpose in mind.
Econometrics is somewhat older and dates to early 20th century with the term in modern sense coined by Norway economist Ragnar Frisch in 1926, which he defined as follows. “Intermediate between mathematics, statistics, and economics, we find a new discipline which, for lack of a better name, may be called econometrics. Econometrics has as its aim to subject abstract laws of theoretical political economy or “pure” economics to experimental and numerical verification, and thus to turn pure economics, as far as possible, into a science in the strict sense of the word.” (Frisch 1926/1971: Sur un problème d’économie pure [On a problem in pure economics]) (link below).
ML is much younger, and it brunched away from Artificial Intelligence research in 1980-90s when a group of scientists, including Jeoffrey Hinton (whose classic course on neural networks you can now find on Coursera) gave up on logical programming and ventured into probabilistic methods. Simultaneously, the goal of research shifted from achieving Artificial Intelligence to tackling mundane real-life problems, which became plenty with the rise of Internet (classifying text and images, sorting spam emails, clustering and retrieving data etc.). Ironically, abandoning attempts to build AI was exactly what brought forward actual AI research and created preconditions for current boom of AI-based technologies.
And here comes the key difference in tradition and approaches of econometrics and ML. Econometrics was conceived and still develops as a tool to study the world while ML is a complex of methods to solve real life problems. Put it another way, econometrics is focused on knowledge discovery, ML is focused on predictions. In an article “Machine Learning: an Applied Econometric Approach” (2017 – link below) Mullainathan and Spiess argue that ML is all about mapping variable X to variable Y while econometrics is about parameter estimation. This is exactly how I see it too.
That said, the dominating economic understanding of ML methods is still rather narrow, and the set of method described by Mullainathan and Spiess is very basic. There is many more to ML than just supervised learning. There is a lot of other accumulated knowledge in ML that can now be transferred to tackling economic problems. For example, my experience suggests that economists should probably pay more attention to unsupervised learning, generative models and auto-encoders. There is also way around poor model interpretability, which is often sighted as the biggest ML drawback. In my following posts (let’s hope I find regular time for that) I will describe how I see potential economic applications for ML, challenges and potential solutions. I also sincerely hope that you will not stay passive and bring your own pieces of knowledge here as well. This is all just too vast for one humble person.
http://www.uio.no/…/undervisnings…/LN-9%20Presentation-1.pdf
https://www.aeaweb.org/articles?id=10.1257/jep.31.2.87
https://en.wikipedia.org/wiki/Timeline_of_machine_learning

Leave a Reply

Your email address will not be published. Required fields are marked *