Modelling and Forecasting Portfolio Inflows : a Comparative Study of Support Vector Regression, Artificial Neural Networks and Structural VAR Models

Abstract

The current study analyses the efficiency of the Support vector regression, artificial neural networks and structural VAR models in terms of in-sample forecasting of portfolio inflows. The study used a time series daily data of portfolio inflows as the dependent variable and real GDP, exchange rate, inflation linked bonds as the independent variables sourced from rand merchant bank and the South African Reserve Bank respectively, and covering the period of 1st March 2004 to 1st February 2016 consisting of a total of 3111 observations. The study assessed non-linearity and non-stationarity prior to modelling the data and based on the results all the variables are nonlinear and non-stationary respectively. The UVAR model employed the SBC criteria in selecting the lag length of the model and the VAR (8) model was selected. Based on the results of the SVAR model 69% of variation in portfolio inflows are explained by the shocks of pull factors (real GDP and inflation linked bonds) and the results are in line with the findings of Egly et al. (2010) who employed VAR model and only 9% is explained by the shocks of push factor (exchange rate) respectively. Furthermore, it is shown by the results that pull factors are the key drivers of portfolio inflows into South Africa. In evaluating model performance, the following error measures are used: Mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), mean absolute squared error (MASE) and root mean scaled log error (RMSLE). The overall results show that support vector regression (SVR) model outperformed competing model(s) as it had the smallest measurement error. The results obtained however can be improved by applying the model to the hybrid technique to improve forecasting accuracy.