Parameter sensitivity of support vector regression and neural networks for forecasting
Sven F. Crone, Stefan Lessmann, Swantje Pietsch · 2006
Support Vector Regression (SVR) and artificial Neural Networks (NN) promise attractive features for time series forecasting. Despite their attractive theoretical properties, limited empirical studies using small or unbalanced parameter setups yield inconsistent results regarding their empirical accuracy. This paper investigates the accuracy of different configurations of NN and SVR parameters, paying particular attention to the common SVR kernels of polynomial, radial basis functions, sigmoid and linear functions through an exhaustive empirical comparison. We investigate the forecasting performance of alternative parameter setups with established benchmarks, evaluating all models on 36 artificial time series with archetypical patterns of level, trend, seasonality and trend-seasonality. As a result, we find that SVR and NN outperform statistical methods on particular time series patterns. Forecasting performance of SVR and NN is impacted by choice of parameters, indicating NN and SVR with the RBF kernels as robust choices on most time series forecasting problems.