Unilateral boundary time series forecasting.

Area of Science:

• Data Science
• Machine Learning

Background:

• Traditional time series forecasting models struggle with unilateral boundary conditions, leading to systematic over or underestimation.
• Existing methods lack specialized frameworks to handle skewed datasets effectively.

Purpose of the Study:

• To develop a novel framework for accurate time series forecasting under unilateral boundary conditions.
• To address underestimation biases in skewed datasets and improve forecast precision.

Main Methods:

• Introduction of unilateral mean square error (UMSE), an asymmetric loss function.
• Implementation of a dual model structure to process underestimated and accurately estimated data separately.
• Utilization of feature reconstruction to recapture obscured data dynamics.

Main Results:

• Demonstrated superior accuracy and robustness of the proposed method using LightGBM and GRU models across diverse datasets.
• Validated the effectiveness of UMSE and the dual model structure in handling underestimation biases.
• Showcased significant improvements over traditional models and existing methods.

Conclusions:

• The novel approach provides accurate and robust time series forecasting for skewed datasets.
• The method is model-independent and broadly applicable across various industries.
• This research paves the way for advanced analytical models in critical forecasting applications.