Univariate time series prediction leveraging linear and nonlinear patterns

Büyükşahin, Ümit Çavuş
Making forecasts from time series data has become increasingly important with the increase of data collection capabilities.The forecasting in time series is mostly done in a univariate setting where historical data in the series itself is used for further time steps. Forecasting in univariate time series is a challenging task due to unpredictable variations in the historic data. The objective of this thesis is to develop prediction methods that exhibit better performance results by alleviating the challenges in univariate time series. The methods we developed for this purpose can be divided into three groups. This thesis first proposes a novel Autoregressive Integrated Moving Average (ARIMA)-Artificial Neural Network(ANN) hybrid method that works in a more general framework. This hybrid method is then expanded to feature-based hybrid approach where statistical and structural features of time series are taken into account. Experimental results show that strategies for decomposing the original data and for combining linear and nonlinear models throughout the hybridization process are key factors in the forecasting performance of the methods. Secondly, by incorporating these findings, this thesis proposes two methods that use Empirical Mode Decomposition (EMD) technique which generates more predictable components. These methods are then enhanced with method selection algorithm which determine the most suitable method for each component. The performed experiments show that our hybrid method with EMD can be an effective way to improve predictive accuracy. These experiments additionally show that having less fluctuations in already stationary time series data leads to more accurate results in forecasting. With the motivation of these consequences, this thesis thirdly proposes another novel method which recursively employs EMD technique for those fast fluctuating components until it achieves more regular and easy-to-predict sub-components. The experiments demonstrate that the recursive algorithm outperforms the previously developed and examined methods. At the end of the thesis, the methods we developed step by step are applied for the prediction of hydropower production data and the results are compared. As a result, the Recursive EMD-based method also produces more accurate predictions for the hydropower data set.


Time Series Forecasting Using Empirical Mode Decomposition and Hybrid Method
Büyükşahin, Ümit Çavuş; Ertekin Bolelli, Şeyda (2018-07-09)
Recently, various applications produce large amount of time series data. In these domains, accurately forecasting time series has been getting important for decision makers. autoregressive integrated moving average (ARIMA) as a linear method and Artificial Neural Networks (ANNs) as a nonlinear method have been widely used to forecast time series. However, many theoretical and empirical studies showed that assembling of those two approaches in hybrid methods can be efficient to improve forecasting performanc...
Prediction Model Selection with Frequency Check on Decomposed Time Series
Büyükşahin, Ümit Çavuş; Ertekin Bolelli, Şeyda (2019-08-22)
High prediction accuracies at time series modeling and forecasting is of the utmost importance for a variety of application domains. Various time series prediction methods exist that use linear and nonlinear models separately or combination of both. These methods highly increase prediction performance results when they are applied on a many number of stationary components obtained by more sophisticated decomposition techniques. Although these stationary components are easily predictable, they each have diff...
Probabilistic Forecasting of Multiple Time Series with Single Recurrent Neural Network
TOPALLAR, SARP TUĞBERK; Yozgatlıgil, Ceylan; Department of Scientific Computing (2022-9-20)
Time series forecasting can be summarized as predicting the future values of a sequence indexed by timestamps based on the past records of that sequence. Optimal or near-optimal resource allocation requires accurate predictions into the future. The study presents investigation performed on both classical methods and more contemporary methods from the literature. The classical methods studied are Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing (ETS) and Seasonal-Trend decomposition us...
A Computational approach to detect inhomogeneities in time series data
Yazıcı, Ceyda; Yozgatlıgil, Ceylan; Batmaz, İnci; Department of Statistics (2017)
Detection of possible inhomogeneity within a series is an important problem in time series data. There are many sources from which inhomogeneity can be originated such as mean shift, variance and trend change, gradual change, or sudden decrease or increase in time series. Since time series has many application areas, the detection of changepoints should be investigated before conducting any analysis. Available methods have certain drawbacks that may lead to unreliable inferences. These include the need of i...
Performance of ensemble forecasting tools for analysis Turkish consumer price index
Aydemir, Petek; Yozgatlıgil, Ceylan; Department of Statistics (2018)
Major challenge in time series analysis is to get reasonably accurate forecasts of the future data from the analysis of the previous records. Accurate forecasting of inflation has great importance in the market economies, the policymakers and the monetary system since the inflation rate determines the cost and standard of living. Also, it affects the decision on investments. In Turkey, the inflation rate is measured by the consumer price index (CPI). There exist many methods to predict the future values of ...
Citation Formats
Ü. Ç. Büyükşahin, “Univariate time series prediction leveraging linear and nonlinear patterns,” Ph.D. - Doctoral Program, Middle East Technical University, 2020.