Machine Learning:An Applied Econometric Approach

Author:Sendhil Mullainathan and Jann Spiess

---

文章目录

• Machine Learning:An Applied Econometric Approach
• 1.Introduction
• 2.How Machine Learning Works
• • 2.1 From Linear Least-Squares to Regression Trees
  • 2.2 The Secret Sauce
  • 2.3 Econometric Guidance（计量经济学的指导）
  • 2.4 Quantifying Predictive Performance
  • 2.5 What Do We (Not) Learn from Machine Learning Output?
  • 2.6 Recovering Structure:Estimation ( β ^ \hat{\beta} β^​) vs Prediction( y ^ \hat{y} y^​)
• 3.How Machine Learning Can Be Applied
• • 3.1 New Data
  • 3.2 Prediction in the Service of Estimation
  • 3.3 Prediction in Policy
  • 3.4 Testing Theory
• 4.Conclusion

---

1.Introduction

机器学习围绕着“预测”的问题；而许多经济学应用则围绕着“参数估计”的问题。机器学习属于预测$\hat{y}$的工具箱的部分，而不是在更熟悉的$\hat{\beta}$的划分下。

empirical economists:实证经济学家
revolve around:围绕着
put succinctly:简而言之
10-K filings: 10-K 适用于美国上市公司。在每个财政年度末后的90天之内（拥有75million美元资产的公司必须在60天之内），公司要向 US Securities and Exchange Commission （SEC）递交10-K表格，内容包括公司历史、结构、股票状况及盈利等情况。

将机器学习算法应用于经济学：第一类，使用新型的数据去解决传统的问题。第二类，估计参数 β \beta β的推断过程中，实际隐含着一个预测任务。（在估计异质性处理效应时）。第三类是直接的政策应用（policy application)

---

提示：以下是本篇文章正文内容，下面案例可供参考

2.How Machine Learning Works

本文只讨论有监督的机器学习算法。由表1可知，最后一行的集成方法（把回归树、LASSO、随机森林）的预测结果取平均，得到的预测结果最好。 R 2 R^2 R2接近46%。

2.1 From Linear Least-Squares to Regression Trees

普通最小二乘法需要手动选择交互变量，而机器学习可以自动搜索交互效应。例如：回归树。
Trees are thus a highly interactive function class. （树是一个高度交互的函数类）

2.2 The Secret Sauce

机器学习吸引人的一点在于它的高维性，灵活的函数形式允许我们去拟合各种各样结构的数据。但是机器学习算法通常有一个与之联系的正则化方法。

2.3 Econometric Guidance（计量经济学的指导）

（1）First, this approach involves choosing the functions we fit and
how we regularize them
（2）Practically, one must decide how to encode and transform the underlying variables. Economic theory and content expertise play a crucial role in guiding where the algorithm looks for structure first.
(3)A final set of choices revolves around the tuning procedure.
Summarize:Design choices must be made about function classes, regularizers, feature representations, and tuning procedures: there are no definitive and universal answers available. This leaves many opportunities for econometric research.

2.4 Quantifying Predictive Performance

计量经济学理论具有着双重角色。一方面，可以指导设计选择（例如“折”的数量或者函数种类）；另一方面，可以使得我们对于拟合效果做出估计。

2.5 What Do We (Not) Learn from Machine Learning Output?

这一节没太看懂

2.6 Recovering Structure:Estimation ( β ^ \hat{\beta} β^​) vs Prediction( y ^ \hat{y} y^​)

A key area of future research
in econometrics and machine learning is to make sense of the estimated prediction
function without making strong assumptions about the underlying true world.
计量经济学和机器学习未来研究的一个关键领域是，在不对潜在的真实世界做出强有力的假设的情况下，理解估计的预测函数。

3.How Machine Learning Can Be Applied

机器学习主要是解决预测问题，即( y ^ \hat{y} y^​) 问题。

3.1 New Data

将以前从未考虑过的非传统数据，例如语言数据、图像数据等，那些too high-dimensional的数据考虑进来，放进回归当中。
举例，卫星图像、线上公告（online posts）。

3.2 Prediction in the Service of Estimation

Chernozhukov, Chetverikov, Demirer, Duflo, Hansen, and Newey (2016) take care of high-dimensional controls in treatment effect estimation by solving two simultaneous prediction problems, one in the outcome and one in the treatment equation.
在处理效应估计中考虑高维控制，同时解决了两个预测问题。
(1)Consider the problem of verifying balance between treatment and control groups (such as when there is attrition). Or consider the seemingly different problem of testing for effects on many outcomes. Both can be viewed as prediction problems.
(2)Estimating heterogeneous treatment effects can also be viewed as a prediction problem 估计异质性处理效应也能够被视作一个预测问题

3.3 Prediction in Policy

3.4 Testing Theory

有监督学习方法能够被应用来检测理论。

4.Conclusion

有启发的句子：
For empiricists, these theory- and data-driven modes of analysis have always coexisted. Many estimation approaches have been (often by necessity) based on top-down, theory-driven, deductive reasoning. At the same time, other approaches have aimed to simply let the data speak. Machine learning provides a powerful tool to hear, more clearly than ever, what the data have to say.

机器学习工具可能也会增加我们工作的范围——不仅是应用新的数据，同时也是使我们聚焦在新的问题上面。