Using data from World Bank, we prepared a data frame with countries' GDP, distance from China, if they have a common border with China, imports from China, exports to China and total trade for years from 2008 to 2018. We classified years and countries according to being a member of BRI or not. This data is the foundation of the famous Gravity model of trade, which goes back all the way to 1954 and introduced by the economist Walter Isard. This model is a modification of the Newtonian gravity formula, Force = G * (Mass_1 * Mass_2) / distance. Surprisingly or not, when we replace mass with economic size, and distance with geographic distance we get very good predictors of actual trade volumes. With a former student of mine, what we did is to test the effectiveness of being a member of BRI using neural networks with gravity model data.
Neural networks are a relatively new area where it is possible to test relationships and make predictions using large datasets. The name comes from an analogy to human neural system, though it's similarities are vastly exaggerated. We are not going to talk about neural networks in detail here but there are lots of resources about how they work on the internet. We used a neural network with one hidden layer with 20 nodes. Neural networks work by training existing data and comparing it with actual (true) values and consequently producing parameters that identify the model. These parameters, similar to coefficients in a linear regression model, are then used for testing and making predictions.
We divided our data set into BRI-member countries and non-BRI-member countries. We then trained our model using BRI-member country data only and produced a parameter set predicting trade volume, accordingly. Our theory is: if being BRI-member country would make a difference, then, when applied to non-BRI member country data, these parameters would predict higher than actual trade volume. In other words: Train the model with BRI-member countries and apply it to non-BRI-member countries for trade volume predictions. If being a BRI-member country would make a difference, you should get higher predicted trade volume than actual trade volume. Which is indeed our result from running the model. Our conclusion is, yes being a BRI-member country makes a difference in improving trade. We couldn't get a 100% vindication rate but (454/496) * 100 = 92%, which is pretty high anyways.
We uploaded the data set and the code to a repository at GitHub which is available to public, for those who would like to run the code on their own. Also, more information is available in commented lines in the code. Feel free to reach us through this post for more information. Here is the link to the GitHub repository.