Comparative Analysis of Three Techniques for Predictions in Time Series Having Repetitive Patterns

Abstract

Modelling nonlinear patterns is possible through using regression (curve fitting) methods. However, they can be modelled by linear regression (LR) methods, too. This kind of modelling is usually used to depict and study trends and it is not used for prediction purposes. Our goal is to study the applicability and accuracy of piecewise linear regression in predicting a target variable in different time spans (where a pattern is being repeated). Using moving average, we identified the split points and then tested our approach on a real world case study. The dataset of the amount of recycling material in Blue Carts in Calgary (including more than 31,000 records) was taken as a case study for evaluating the performance of the proposed approach. Root mean square error (RMSE) and Spearman rho were used to evaluate and prove the applicability of this prediction approach and evaluate its performance. A comparison between the performances of Support Vector Machine (SVM), Neural Networks (NN), and the proposed LR-based prediction approach is also presented. The results show that the proposed approach works very well for such prediction purposes. It outperforms SVM and is a powerful competitor for NN.