Scribd
Upload
18 views

Course0 Introduction

Uploaded by

hishamhaydar
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
18 views

Course0 Introduction

Uploaded by

hishamhaydar
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 14

Machine Learning for Econometrics

Introduction

Christophe Gaillac

Autumn 2024
Motivation

Econometrics and machine learning (ML) share many statistical


tools, but (usually) have different 1) purposes, 2) approaches to
building models.
Econometrics, first and foremost, aims to quantify a precisely
defined effect, with a focus on statistical inference:
what is the impact of a minimum wage increase on employment?
Are there peer effects among groups of students?
What is the average wage gap between women and men?
Empirical economics has focused on measuring causal effects,
stressing the identification strategy.
The goal of ML is to build a model that allows one to obtain the
best possible predictive performance for a given problem:
computational constraint when calling the model (runtime or
inference time performance)
focus on algorithms rather than models

Introduction 2 / 12
Motivation

Econometrics models are preferably simpler, with an emphasis on


linearity, and their structure is usually motivated by a theory of the
underlying causal relationship or of individual behavior
In ML, prediction performance is usually the only criterion for
selecting a model.

Some nuances:
several scandals involving the use of algorithms have highlighted the
need to be able to explain their predictions (so called explainability ),
to study and correct their biases (see the fairness and bias
mitigation literature)
“p-hacking” (i.e., the practice of embarking on a specification search
to obtain significant results), or more generally the replicability crisis
in empirical economics can be seen as a problem of overfitting.

Introduction 3 / 12
Outline

1 What is this course about?

2 Additional references

Introduction 4 / 12
High dimension and variable selection

Empirical economics involves crucial choices about the model: may


cast doubt on the credibility of the results.

First focus on high-dimensional methods, which can handle a


large number of covariates and/or instrumental variables, and on
certain machine learning techniques with the purpose of performing
causal inference in mind.

Introduction 5 / 12
High dimension and variable selection

Example: using the Neyman-Rubin potential outcome model.


Yi (0) is the potential outcome for individual i if not treated and
Yi (1) is the potential outcome if treated. We only observe the state
of treatment Di ∈ {0, 1} and the realized outcome Yi defined by:

Yi (0) if Di = 0,
Yi = Yi (Di ) =
Yi (1) if Di = 1.

One interesting quantity is the average treatment effect

τ0 := E[Yi (1) − Yi (0)],

representing the average impact of the intervention.

Introduction 6 / 12
High dimension and variable selection

1 If treatment assignment is random when conditioning on observables


(i.e., assuming that E[ε i |Di , Xi ] = 0 in the model below) and there
are only a limited number of significant covariates (sparsity) provides
the tools to estimate τ0 in the model:
′
Yi = Di τ0 + Xi β 0 + ε i , with E[ε i ] = 0 and E[ε i |Di , Xi ] = 0,

where Xi is a vector of p exogenous control variables, p being


potentially larger than the number of observations.
2 Generalization to other models, ML predictors
3 Generalization to E[ε i |Di , Xi ] ̸= 0 and many instruments.

Introduction 7 / 12
Estimation of heterogeneous effects
The average treatment effect (τ0 above) does not describe the
heterogeneity of responses to an intervention – some people may
benefit greatly from it, while others may not be affected or may
even be worse off.
Focus on a more complex parameter of interest, which is the
average treatment effect conditional on certain (observed) variables
τ : x 7→ E[Yi (1) − Yi (0)|Xi = x ].

Objective: inference on the function τ (·), i.e., to test the


significance of the effect conditional on covariates taking the value x .
Then lower our requirements to make inference only on certain
characteristics of the conditional average treatment effect, while
ensuring better theoretical guarantees
Then focus on optimal policy learning, which presents the tools
for estimating optimal policies in the context of randomized
experiments.
Introduction 8 / 12
Aggregate data and macroeconomic forecasting

Tools to handle data that has a temporal structure, often taking a


more aggregated form in economics.

High-dimensional estimation methods in a context where the


data is not identically distributed and potentially has heavy tails,
as is the case with macroeconomic and financial data.
Adapt the tools developed in the previous chapters to that of
predicting macroeconomic variables, in the context where one wishes
to select without prior from a large number of explanatory variables.
Talk about the limitations of sparse methods in this context are also
underlined and links to factor models, which are “dense” models, are
made.

Introduction 9 / 12
Text data

Short introduction to the basic tools for analysis of text data

First part on tools to extract a numerical representation of texts, as


well as to model the language.

Second part on the fundamental concepts of modern natural


language processing (NLP): word embeddings. Then, we will see
generalizations of the concept of embeddings beyond textual data,
still experimental in empirical economics, but identified as promising
for the future.

Introduction 10 / 12
Outline

1 What is this course about?

2 Additional references

Introduction 11 / 12
Resources

General econometrics
1 Wooldridge, J. M. (2002). Econometric analysis of cross section and
panel data. Cambridge and London: MIT Press
2 Hansen, B. E. (2022). Econometrics. Princeton University Press.

Causal inference
1 Angrist, J. and J.-S. Pischke (2009). Mostly Harmless Econometrics:
An Empiricist’s Companion (1st ed.). Princeton University Press.
2 Chernozhukov, V., C. Hansen, N. Kallus, M. Spindler, and V.
Syrgkanis (2024). Applied causal inference powered by ML and AI.

Machine learning methods, Neural Networks:


1 Hastie, T., R. Tibshirani, and J. Friedman (2009). The Elements of
Statistical Learning: Data mining, Inference and Prediction. Springer
2 Goodfellow, I. J., Y. Bengio, and A. Courville (2016). Deep Learning.
Cambridge, MA, USA: MIT Press. http://www.deeplearningbook.org.

Introduction 12 / 12
Resources

Our book with Jérémy L’Hour (CFM):


1 Gaillac, C. and J. L’Hour (2023). Machine Learning pour
l’économétrie. Economica
2 Gaillac, C. and J. L’Hour (June 2025). Machine Learning for
Econometrics. Oxford University Press

A GitHub repository is available at the address


https://github.com/jeremylhour/ml econometrie.
It contains scripts in R and Python, which reproduce some of the
applications.

Introduction 13 / 12
Appendix

Introduction 14 / 12

You might also like