Clustered Regression Models for Analysis and Prediction of Foreign Direct Investment Inflows

Abstract: Foreign Direct Investment (FDI) is one of many economy components that has been found to be attractive to many countries as one alternative of private capital inflow. It has a crucial role in achieving rapid economic growth, especially in developing countries. In fact, most of the direct investment takes place among developed countries. FDI involves two parties, which are the host country and the investing country and FDI benefits both of them. There are particular characteristics of the host country that are crucial for the investing country when determining in which host country they do the investment. It is therefore important to investigate the potential FDI determinants and their relationships to predict the future trend of FDI inflows to the host country. However the availability of FDI series is limited for each country. Therefore we propose a novel approach by coupling clustering methods with random forest regression or Bayesian Model Averaging (BMA) to tackle these problems. Random forest regression is a promising approach for prediction of financial time series but has never been used in predicting FDI inflows. BMA is another approaches in economic forecasting, which is gaining more attention. Using various characteristics of countries from various regions and the FDI determinants that have been identified in the literature on FDI in both developing and developed countries in Asia, Africa, Latin America and Europe, the analysis is conducted in one framework that has been performed separataly between developed and developing countries or by the region. In the first stage of the proposed solution, the individual countries are clustered into a smaller sets of clusters to reduce the uncertainty and to improve the forecasting performance compared to estimating the individual time series models for each country separately. Then for each cluster, random forest regression and BMA solve the proposed regression equation, in which each cluster the model parameters are restricted to be the same to gain estimation precision. To evaluate the performance of the proposed solution, the prediction error for each model with and without clustering will be compared. The results show an improvement as much as 70.31% with the proposed solution.