Sufficient Forecasting Using Factor Models.

Area of Science:

• Statistics
• Econometrics
• Machine Learning

Background:

• Traditional forecasting methods struggle with high-dimensional predictor sets and potential nonlinearities.
• Dimensionality reduction is crucial for effective modeling in such scenarios.

Purpose of the Study:

• To develop a novel forecasting method, sufficient forecasting, for time series with numerous predictors and potential nonlinear effects.
• To extend sufficient dimension reduction techniques to high-dimensional settings.

Main Methods:

• Dimensionality reduction using high-dimensional (approximate) factor models via principal component analysis.
• Development of sufficient forecasting utilizing extracted factors to identify sufficient predictive indices.
• Application of projected principal component analysis for enhanced factor inference in semi-parametric models.

Main Results:

• Sufficient forecasting provides a set of sufficient predictive indices, enhancing predictive power from high-dimensional predictors.
• The method correctly estimates projection indices of underlying factors, even with nonparametric forecasting functions.
• Asymptotic properties for central subspace and sufficient predictive index estimates are derived.
• The method accommodates more predictors than observations.

Conclusions:

• Sufficient forecasting offers improved accuracy over linear forecasting for time series prediction with high-dimensional predictors.
• The proposed method effectively condenses cross-sectional information through factor models.
• The connection between sufficient forecasting and deep learning architectures is established.