Evolutionary support vector machines for time series forecasting

• Author(s): Cortez, Paulo, 1971- ; Peralta Donate, Juan
• Date: 2012
• Persistent ID: http://hdl.handle.net/1822/21407
• Origin: RepositóriUM - Universidade do Minho
• Subject(s): Evolutionary computation; Support vector machines; Time series; Forecasting
• Document Type: Conference Object
• Language: English

Description

Abstract. Time Series Forecasting (TSF) uses past patterns of an event in order to predict its future values and is a key tool to support decision making. In the last decades, Computational Intelligence (CI) techniques, such as Artificial Neural Networks (ANN) and more recently Support Vector Machines (SVM), have been proposed for TSF. The accuracy of the best CI model is affected by both the selection of input time lags and the model’s hyperparameters. In this work, we propose a novel Evolutionary SVM (ESVM) approach for TSF based on the Estimation Distribution Algorithm to search for the best number of inputs and SVM hyperparameters. Several experiments were held, using a set of six time series from distinct real-world domains. Overall, the proposed ESVM is competitive when compared with an Evolutionary ANN (EANN) and the popular ARIMA methodology, while consuming less computational effort when compared with EANN.