Wavelet-SVR: a novel wavelet-based support vector regression algorithm for time series forecasting

Autoregressive integrated moving average (ARIMA) and its constituent models are generally used in modelling linear pattern. Among the nonparametric models, artificial intelligence (AI) tools such as artificial neural networks (ANNs), random forests (RFs), and support vector regression (SVR) are evolving within the area of forecasting. However, for complex data, noise makes training patterns difficult for these models. To handle the complex and noisy data, the present study proposes an efficient hybrid algorithm, namely Wavelet-SVR, which uses a denoising technique, and an SVR model for time series forecasting. First, wavelet transformation is used to attenuate background noise, and then the SVR model is applied to each denoised series. Finally, the ensemble technique is applied to obtain the final prediction. The proposed Wavelet-SVR algorithm is compared with available benchmark models via four error measures. For empirical evaluation, the wholesale price of potatoes in major markets in India has been considered, as these datasets are habitually nonlinear, nonstationary, nonnormal, and heteroscedastic in nature. This study empirically concluded that the proposed Wavelet-SVR model outperformed the other models in terms of prediction efficiency. A user-friendly R package has been developed and made available on CRAN to facilitate the easy implementation of these proposed algorithms.