The Determinants of Indonesia Export Performance: Gravity Model and Random Forest Model

In the last six decades, economists have used the gravity model to predict one country’s exports and their determinants. However, this method is sensitive to a large dataset, a heavy missing value dataset, and an unbalanced dataset. This method needs to normalise the data heavily and cherry-picked a subset of data to make the model work. Recently, a new method from the computer science field called machine learning was implemented by applied economists in their respective problems. One machine learning method to predict a dependent variable from its independent variable is a random forest. This method does not need to normalise, can process raw data, and handles large and sparse datasets. We are experimenting with this new method to handle the large dataset we collect from various sources. Some variables in our dataset have been proven to affect exports, but another researcher has not yet explored others. This paper uses a random forest to predict exports and their determinants. We achieve the result of a $\mathbf{9 2 . 5 \%} \mathbf{R}^{2}$ after some experiments. This is higher than the result from the gravity model that has $63.8 \% \mathbf{R}^{2}$ with Fixed Effect and $\mathbf{8 5 . 2 \%} \mathbf{R}^{\mathbf{2}}$ with Pooled Least Square. With random forest, we can also list the variables that determine Indonesia’s export to the target country, with the degree of importance of each variable.