계량경제학과 머신러닝: 토빗, 헷킷과의 예측력 비교

초록

This paper compares the predictive power of Tobit and Heckit, which are used for the parameter estimation when the dependent variable has limited characteristic, respectively against machine learning methodologies (MLM). The methodologies refer to SVR(Support Vector Regression), RF(Random Forest), GBRT(Gradient Boost Regression Tree). The data employed in this paper is one typically used for explaining Tobit and Heckit in well-known econometric textbooks. The results are as follow. First, in the comparison of Tobit vs. MLM, RF is the best in the in-sample prediction, but Tobit is the best in the out-of-sample prediction with a slight difference. Second, in the comparison of Heckit vs. MLM, RF is the best in the in-sample prediction, but GBRT in RMSE and Heckit in MAE are the most excellent in the out-of-sample prediction with a slight difference. That is, Tobit and Heckit mainly focus on the accurate estimation of parameters, and MLM focus on forecasting power, but it can be seen that the former method is not significantly different from MLM in terms of forecasting power. The implication of this result is that the comparison of the forecasting power between the traditional econometric techniques and MLM can help to learn the two methods.

키워드