Modelling and forecasting foreign direct investment: A comparative application of machine learning based evolutionary algorithms hybrid models

Abstract

The study aims to determine whether genetic algorithms can improve the forecasting accuracy of machine learning models in modelling and forecasting foreign direct investment (FDI). The use of evolutionary algorithms for FDI modelling and forecasting is relatively unexplored. The research therefore employs benchmark models (ANN, SVR, and LSSVR) to assess the forecasting performance of hybrid models (ANN-GA and SVR-GA) in modelling and forecasting FDI time series data. Monthly time series data of FDI (dependent variable) and GDP, inflation rate, and exchange rate (explanatory variables) were collected from World Bank and South African Reserve Bank spanning January 1970 to June 2019, with 594 observations.

The nonlinearity tests conducted in the study confirmed that the variables under considerationexhibit nonlinear characteristics. The nonlinearity tests considered were Brock, Dechert and Scheinkman (BDS) test, Cumulative Sum (CUSUM) test, Keenan’s test, White and Terasvirta neural network test, Tsay test and McLeod-Li test. In traditional machine learning models, artificial neural networks (ANN) outperformed support vector regression (SVR) and least squares support vector regression (LSSVR), showing lower error measures. It was further found that the support vector regression (SVR) was the second-best performing model and the worst being the least squares support vector regression (LSSVR). Among hybrid models, support vector regression based genetic algorithm (SVR-GA) performed better than artificial neural network based genetic algorithm (ANN-GA). However, overall, ANN proved to be the best performing model in terms of error measurements (MSE, RMSE, MAE, MAPE, and MASE) with the lowest error values reported. Genetic algorithm (GA) did not significantly improve the forecasting performance of the hybrid models, which still fell short compared to traditional models. The decision of the error measurements was supported by Diebold Mariano test, concluding that artificial neural network (ANN) performed best in forecasting accuracy, followed by support vector regression (SVR) and least squares support vector regression (LSSVR). Both ANN and SVR exhibited similar forecasting accuracy for testing predictive power. Among hybrid models, ANNGA showed higher predictive accuracy than SVR-GA. Overall, machine learning models proved effective in modelling and forecasting foreign direct investment. The study recommends exploring other evolutionary algorithm(s) methods to optimise hyperparameters and suggests employing a wider range of machine learning models for FDI forecasting