Probabilistic-Based Forecasting Method For Time Series Datasets

Abstract

In this paper, a new probabilistic technique (a variant of Multiple Model Particle Filter-MMPF) will be used to predict time-series datasets. At first, the reliable performance of our method is proved using a virtual random scenario containing sixty successive days; a large difference between the predicted states and the real corresponding values arises on the second, third, and fourth day. The predicted states that are determined by using our method converge rapidly towards the real values while a classical linear model exhibits a large amount of divergence if used alone here. Then, the performance of our approach is compared with some other techniques that were already applied to the same time-series datasets: IEX (Istanbul Stock Exchange Index), TAIEX (Taiwan Stock Exchange), and ABC (The Australian Beer Consumption). The performance evaluation metrics that are utilized here are the correlation coefficient, the mean absolute percentage error, and the root mean squared error.

Keywords

• forecasting; time series dataset
• mmpf
• evaluation metrics
• tahmin; zaman serisi veri kümesi; çmpf; değerlendirme metrikleri