Periodic Neural Networks for Multivariate Time Series Analysis and Forecasting

Abstract

Designing systems that make accurate forecasts based on time dependent data is always a challenging and significant task. In this regard, a number of statistics and neural network-based models have been proposed for analyzing and forecasting time series datasets. In this paper, we propose a novel machine learning model for handling and predicting multivariate time series data. In our proposed model we focus on supervised learning technique in which (1) some features of time series dataset exhibit periodic behaviour and (2) time t is considered as an input feature. Due to periodic nature of multivariate time series datasets, our model is a simple neural network where the inputs to the single output source are assumed to be in the form A sin(Bt + C)x as opposed to the standard form inputs Ax + B. We train our proposed model on various datasets and compare our model’s performance with standard well-known models used in forecasting multivariate time series datasets. Our results show that our proposed model often outperforms other exiting models in terms of prediction accuracy. Moreover, our results show that the proposed model can handle time series data with missing values and also input data-values that are non-equidistant. We hope that the proposed model will be useful in fostering future research on designing accurate forecasting algorithms.