Optimized ARIMA-ANN hybrid model for time series analysis


Erturan M. B., Merdivenci F.

JOURNAL OF THE FACULTY OF ENGINEERING AND ARCHITECTURE OF GAZI UNIVERSITY, vol.37, no.2, pp.1019-1032, 2022 (SCI-Expanded)

  • Publication Type: Article / Article
  • Volume: 37 Issue: 2
  • Publication Date: 2022
  • Journal Name: JOURNAL OF THE FACULTY OF ENGINEERING AND ARCHITECTURE OF GAZI UNIVERSITY
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Art Source, Compendex, TR DİZİN (ULAKBİM)
  • Page Numbers: pp.1019-1032
  • Akdeniz University Affiliated: Yes

Abstract

Purpose: The purpose of this study is to present a novel ARIMA-ANN hybrid model for time series analysis. Proposed optimized ARIMA-ANN (OptAA) hybrid model is applied to three well-known time series data with varying forecasting horizons. For determination of the forecasting performance, results are compared with the results of other models. Theory and Methods: Proposed model uses a least squares optimization of ARIMA and ANN models of the time series data to decompose it into linear and nonlinear components. After the first decomposition, error series of the linear part is transferred to the nonlinear component to revise the nonlinear part, which is then remodeled with ANN. The sum of the ARIMA model forecast of the linear part and ANN forecast of the revised nonlinear part is the final forecast of the hybrid model. Three time series data, Wolf’s sunspot, Canadian lynx and GBP/USD exchange rate are used for forecasting performance comparison purposes. Proposed hybrid model’s forecasting performance is compared with four major ARIMA-ANN hybrid models, ARIMA, ANN and random walk model. Results: Obtained results show that OptAA hybrid model is a very powerful methodology for time series forecasting. Especially for short term forecasting horizons proposed hybrid model shows better performance than other compared models. Conclusion: OptAA hybrid model is open for further research. Testing the model with different neural network parameters such that learning algorithm, network architecture, activation functions etc. is possible. Also applying the model to different time series and forecasting horizons helps to improve the generalization of the model.