A hybrid forecasting algorithm based on SVR and wavelet decomposition

Abstract

We present a forecasting algorithm based on support vector regression emphasizing the practical benefits of wavelets for financial time series. We utilize an e ective de-noising algorithm based on wavelets feasible under the assumption that the data is generated by a systematic pattern plus random noise. The learning algorithm focuses solely on the time frequency components, instead of the full time series, leading to a more general approach. Our findings propose how machine learning can be useful for data science applications in combination with signal processing methods. The timefrequency decomposition enables the learning algorithm to solely focus on periodical components that are beneficial to the forecasting power as we drop features with low explanatory power. The proposed integration of feature selection and parameter optimization in a single optimization step enable the proposed algorithm to be scaled for a variety of applications. Applying the algorithm to real life financial data shows wavelet decompositions based on the Daubechie and Coiflet basis functions to deliver the best results for the classification task.