Forecasting intermittent and sparse time series: A unified probabilistic framework via deep renewal processes.

Area of Science:

• Operations Research
• Statistical Modeling
• Time Series Analysis

Background:

• Intermittent demand poses significant challenges in forecasting accuracy.
• Existing methods often struggle to capture complex demand patterns like aging and clustering.

Purpose of the Study:

• To introduce a unified probabilistic forecasting framework for intermittent demand time series.
• To generalize and extend existing intermittent demand forecasting models.
• To incorporate neural network-based models into the framework.

Main Methods:

• Extending model-based methods to discrete-time renewal processes.
• Integrating recurrent neural networks for demand pattern analysis.
• Developing a framework adaptable to continuous-time demand arrivals via temporal point processes.

Main Results:

• The framework successfully incorporates and generalizes existing methods like Croston-type models.
• Neural network integration enhances the modeling of demand arrival patterns.
• Empirical studies confirm the framework's predictive accuracy on standard datasets.

Conclusions:

• The proposed unified framework offers a flexible and powerful approach to intermittent demand forecasting.
• The framework's adaptability to both discrete and continuous time models enhances its practical applicability.
• This research advances the state-of-the-art in probabilistic forecasting for intermittent demand.