A New Approach for Time Series Prediction Using Ensembles of IT2FNN Models with Optimization of Fuzzy Integrators

Abstract

This paper describes a new approach for time series prediction based on using different soft computing techniques, such as neural networks (NNs), type-1 and type-2 fuzzy logic systems and bio-inspired algorithms, where each of these intelligent techniques can provide a variety of features for solving real and complex problems. Therefore, this paper describes the application of ensembles of interval type-2 fuzzy neural network (IT2FNN) models. The IT2FNN uses hybrid learning algorithm techniques from NNs models and fuzzy logic systems. The output of the Ensemble of IT2FNN models needs the integration process to forecast the time series, and we are required to design the fuzzy integrator (FI) to solve this real problem. Genetic algorithms and particle swarm optimization are used for the optimization of the parameter values in the membership functions of the FI. We consider different time series to measure the performance of the proposed model, and these time series are: Mackey–Glass, Mexican Stock Exchange (MSE or BMV), Dow Jones and NASDAQ. The forecasting errors are calculated as follows: mean absolute error, mean square error (MSE), root-mean-square error, mean percentage error and mean absolute percentage error. The best prediction errors are illustrated as follows: 0.00025 for the Mackey–Glass, 0.01012 for the MSE, 0.01307 for the Dow Jones and 0.01171 for the NASDAQ time series. Simulation results are compared using a statistical test and provide evidence of the potential advantages of the proposed approach.