Penalized spline estimation for functional coefficient regression models.

Area of Science:

• Statistics
• Econometrics
• Data Science

Background:

• Functional coefficient regression models capture dynamic relationships by allowing coefficients to vary.
• Traditional methods can struggle with high-dimensional data (curse of dimensionality).
• Existing approaches may lack computational efficiency or explicit model forms for forecasting.

Purpose of the Study:

• To develop and investigate penalized spline (P-spline) methods for functional coefficient regression with dependent observations.
• To provide efficient estimation, inference, and multi-step-ahead forecasting capabilities.
• To explore and compare various smoothing parameter selection techniques.

Main Methods:

• Utilized penalized splines (P-splines) for estimating functional coefficient regression models.
• Employed fixed-knot asymptotics for statistical inference.
• Investigated smoothing parameter selection using modified multi-fold cross-validation (MCV), generalized cross-validation (GCV), empirical bias bandwidth selection (EBBS), and restricted maximum likelihood (REML).

Main Results:

• The P-spline approach demonstrated computational efficiency and stability.
• Established fixed-knot asymptotics facilitate readily available inference, including exact inference for fixed smoothing parameters.
• The method allows for different smoothness levels for individual functional coefficients via distinct penalty parameters.
• Multi-step-ahead forecasting is enabled through explicit model expressions.

Conclusions:

• Penalized splines offer a flexible and efficient framework for functional coefficient regression, particularly with dependent data.
• The proposed methods provide robust estimation, inference, and forecasting capabilities.
• The study validates the approach through simulations and a real-world data application.