Financial time series forecasting: a comprehensive analysis of univariate nonlinear and linear models' performance
Financial markets have always been attractive as a means of increasing one's wealth, and those who make accurate predictions take the prize. Forecasting models such as linear ones are simple to compute, however, they give rough approximations of the underlying relationships in the data, thus, producing poor forecasts. The solution to this issue could be the nonlinear models which try to fit the data and display the relationships with higher accuracy. Previous research seems to prove this statement from the statistician's point of view which might be of little use for an investor. Therefore, the focus of this paper is on the comparison of three types of models (nonlinear: ANN, STAR, and linear: AR) in terms of financial performance. Our research is based on the initial code for GAUSS and papers by Dick van Dijk. The data used is the monthly S&P 500 Index values from 1970 to 2012 provided by the Robert Shiller's website. Forecasting index changes begins at 1995 and ends in 2012 providing up-to-date results for 14 model specifications. The best model proves to be the flexible ANN, beating the linear AR in the majority of cases, leaving the underperforming heavy-parameterized STAR model behind. Thus, it is evident that the more flexible nonlinear models outperform the heavily parameterized ones as well as linear models for the S&P 500 Index. The introduced type of performance evaluation has a more comprehensible application to the financial market analysis.