PROJECT TITLE :

A Hybrid System Based on Dynamic Selection for Time Series Forecasting

ABSTRACT:

Hybrid systems, which combine statistical and machine learning (ML) techniques using residual (error forecasting) modeling, have been highlighted in the literature because of their accuracy and ability to forecast time series with different characteristics. This is due to the fact that hybrid systems combine statistical and ML techniques using residual modeling. The correct modeling of the residuals is a crucial task in these architectures because the residuals may exhibit random fluctuations, complex nonlinear patterns, and heteroscedastic behavior. This makes the modeling of the residuals a crucial task. Since problems like underfitting, overfitting, and misspecification can result in a system with low accuracy or even deteriorate the linear forecast of the time series, the selection, specification, and training of one ML model to forecast the residuals are tasks that are both challenging and expensive to complete. The purpose of this article is to propose a hybrid system that has been given the name dynamic residual forecasting (DReF). This system uses a modified version of the dynamic selection (DS) algorithm to make the following determinations: the most suitable machine learning model to forecast a pattern of the residual series; and whether or not it is a promising candidate to increase the accuracy of the time series forecast derived from the linear combination. Therefore, the purpose of the DReF is to lessen the uncertainty associated with the selection of the ML model and to prevent the time series forecast from deteriorating. In addition to this, the system that has been proposed looks for the parameters of the DS algorithm that are the most appropriate for each data set. The method that is proposed in this article makes use of a pool of five machine learning models that are widely used in the research community. These models are as follows: multilayer perceptron, support vector regression, radial basis function, long short-term memory, and convolutional neural network. The use of ten different well-known time series in an experimental evaluation was carried out. According to the findings, the DReF achieves better results for the vast majority of the data sets when compared to single and hybrid models found in the existing body of academic research.