A grey seasonal least square support vector regression model for time series forecasting

Area of Science:

• Time Series Analysis
• Econometrics
• Machine Learning

Background:

• Seasonality is a prevalent characteristic in real-world time series data.
• Accurate modeling of seasonal variations is crucial for effective forecasting.
• Existing models may not fully capture complex seasonal patterns and nonlinearities.

Purpose of the Study:

• To propose a novel Grey Seasonal Least Square Support Vector Regression (GSLSSVR) model.
• To enhance the realism and accuracy of time series forecasting by incorporating arbitrary seasonality.
• To improve model stability and generalization through a regulation method.

Main Methods:

• Combining dummy variables, Least Square Support Vector Regression (LSSVR), and grey accumulation generation.
• Utilizing the Lagrange multipliers algorithm for parameter estimation.
• Implementing a last block evaluation method for hyperparameter tuning.

Main Results:

• The GSLSSVR model effectively captures seasonal variations in functional forms, variables, and parameters.
• Experimental results on four diverse Chinese seasonal time series validate the model's efficacy.
• The proposed model outperforms SGM(1,1), SFGM(1,1), LSSVR, SARIMA-GARCH, and BPNN models.

Conclusions:

• The GSLSSVR model offers an intuitive and simple framework for handling arbitrary seasonality.
• The model provides a robust approach for analyzing seasonal regulatory measures.
• GSLSSVR is highly recommended for time series with periodic and nonlinear characteristics.