Probabilistic Forecasting With Modified N-BEATS Networks.

Area of Science:

• Machine Learning
• Time Series Analysis
• Deep Learning

Background:

• Univariate time series forecasting often requires probabilistic outputs.
• Existing deep learning models may lack stability or struggle with cumulative forecasts.

Purpose of the Study:

• To modify the N-BEATS architecture for parametric probabilistic time series forecasting.
• To introduce extensions for optimizing forecast stability and jointly forecasting marginal and cumulative values.
• To evaluate the enhanced model's performance in a supply chain context.

Main Methods:

• Modification of the state-of-the-art N-BEATS deep learning architecture.
• Development of extensions to optimize for forecast accuracy and stability.
• Joint optimization of single-period marginal and multiperiod cumulative probabilistic forecasts.
• Empirical evaluation on the M4 monthly dataset.

Main Results:

• The enhanced N-BEATS model provides more stable forecast distributions.
• Minimal loss in forecast accuracy was observed with improved stability.
• The second extension demonstrated improved accuracy for probabilistic cumulative forecasts.
• The model shows utility in supply chain planning.

Conclusions:

• The proposed probabilistic N-BEATS network and its extensions offer significant improvements for time series forecasting.
• The enhancements address forecast stability and cumulative forecasting challenges.
• The model is a valuable tool for practical applications such as supply chain management.