Performance Improvement in Time Series Prediction through PECNET Framework

Machine learning algorithms have garnered con-siderable attention and recognition in the context of addressing time series prediction challenges. However, constructing a ma-chine learning model with optimal architecture and hyperpa-rameters that effectively captures the intricacies inherent in the time series data can present challenges. If the data encompasses multivariable characteristics with chaotic or stochastic proper-ties and has missing parts, the task of constructing an accurate machine learning model with appropriate structure and hy-perparameters becomes further challenging. Addressing the challenge of overfitting, which is a common machine learning problem encountered in such cases where the data exhibits these characteristics, a cascaded neural network, named PECNET (Predictive Error Compensated Wavelet Neural Networks), is a highly favorable alternative to mitigate this issue. PECNET ad-dresses the problem by training separate neural networks in a cascaded manner for each different frequency band of input data features, starting from the input feature most correlated with the target data and applying wavelet transformation to the input data prior to training and utilizing the remaining errors of the prior network as the target labels of the next network. This approach enhances the orthogonality of data characteris-tics across time windows and subsequently reduces the likeli-hood of overfitting as additional networks are added, thereby improving prediction performance. In this study, to tackle the escalation in computational complexity and preempt the occur-rence of implementation errors in time synchronization man-agement, the previously experimentally tested PECNET is transformed into a modular and parametric framework soft-ware, considering the prevalent utilization of off-the-shelf framework software in the majority of artificial intelligence studies. The performance of the developed framework software is initially demonstrated on financial data, and subsequently, within the context of the earthquake prediction project at istan-bul Technical University, it is assessed for the earthquake pre-diction process using chaotic time series data from the Electro-static Rock Stress (ERS) monitoring method.