Time series modeling and forecasting using memetic algorithms for regime-switching models.

Area of Science:

• Time Series Analysis
• Machine Learning
• Computational Statistics

Background:

• The neuro-coefficient smooth transition autoregressive (NCSTAR) model offers interpretable forecasting by integrating fuzzy rule-based systems.
• Traditional NCSTAR model fitting combines grid search with local search algorithms, potentially limiting accuracy.
• Accurate and interpretable forecasting models are crucial for various applications.

Purpose of the Study:

• To propose a novel, more accurate model fitting procedure for the NCSTAR model.
• To enhance the predictive accuracy of NCSTAR models using evolutionary computation.
• To evaluate the performance of the proposed fitting procedure against existing methods.

Main Methods:

• Development of a new model fitting procedure utilizing a memetic algorithm.
• Application of the memetic algorithm to estimate parameters for the NCSTAR model.
• Empirical evaluation on diverse real-world time series datasets from forecasting competitions.

Main Results:

• The memetic algorithm-based fitting procedure significantly improves the accuracy of NCSTAR models.
• The enhanced NCSTAR models demonstrate competitive performance compared to established forecasting models.
• The proposed method offers a statistically founded and accurate approach to model fitting.

Conclusions:

• Memetic algorithms provide a superior alternative for fitting NCSTAR models, enhancing forecasting accuracy.
• The novel fitting procedure makes NCSTAR models more competitive in practical forecasting scenarios.
• This research contributes a more effective tool for building accurate and interpretable time series models.