So obviously this article is going to be written in English, yay.
Since I haven’t write much English recently, my English might be incomprehensible for some part. Let me know in that case on facebook.
Well, I use Machine Learning, which is statistical and algorithmic pure prediction model, to my job in order to predict consumer’s response for ads on the internet.
Machine Learning is based on different idea from Econometrics, it is designed to predict out sample of data set. On the other hands, Econometrics is designed to estimate the causal effect of the variable which you are interested on dependent variable, and such model is not necessary to have predictive ability.
This difference of designs is mainly from the requirement for those model. Economics want to know the theoretical model framework exist in real world, and computer science want to know the outcome of unknown sample.
So what happen if you try to use Machine Learning for Economics?
There is one way that Hal Varian, Chief Economist in Google, stated in his paper, which use the prediction model for using causal inference.
And there is paper in NBER motivated by the above Varian’s paper.
The idea of these methods is simple.
1. divide data set into two parts. The one contains the interesting variable and the other one does not contain the interesting variable. For example, the data that variable of advertisement is 0 and the data which advertisement is equal to one, if you are interested in the effect of advertisement.
2. make a prediction model in the data set that does not contain the interesting variable. In the advertisement example, estimating sales prediction model in the data set where advertisement variable is equal to zero. This model is able to understand as the model to show what is the outcome in the world where there is no advertisement .
3. predict the outcome with the model for the other data set which contain the interesting variable. In the advertisement example, predict the sales for the data set where advertisement is not equal to zero. So results are predicted sales in the case where advertisement is zero.
4. calculate the gap between predicted result and actual data. In other word, predicted sales and actual sales. So if those two data sets are different in advertising, this difference is the effect of advertising.
Well, this is incredible idea. At least for me.
But this article is not about this idea. There is one more idea for using Machine Learning for Economics. So you should check papers I referenced if you wanna know the idea I explained above.
The other way is explained in this paper published on American Economic Review.
This paper states that prediction model is able to support economic policy which usually only concern about causal effect.
So when think about substitution or tax, policy maker want to put incentive in certain action such as buying environmentally friendly goods or take specific surgery in order to maximize social welfare.
However, usually substitution and tax are given to the individual equally, although the effect of such incentive is different among individuals. The above paper shows that the certain surgery for joint, such as knee, have less effect on utility for the person who have low expectancy for survive. Maybe survive is much more important than his/her knee in such case. In such case, the substitution should be allocated to other purpose in order to maximize social welfare.
So the above paper used Machine Learning to predict the survive rate then simulated how much the government can reallocate the substitution from the one who have less probability to survive. Although the author did not mention about what they can do from the saving, it is not so difficult to imagine there is more cost effective option for using such money.
Maybe this is much more simple way to utilize ML into Economics.
Machine Learning is one of the most interesting area in 21th century, and Economics is one of the most interesting area of human history for me. I am really happy that such break through happen in these days.