A new automatic forecasting method based on explainable deep dendritic artificial neural network

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Automated forecasting methods are crucial for practitioners, reducing subjective decisions in model selection and data preprocessing.
• Existing automated methods often rely on complex model/variable selection and hypothesis testing.
• Explainability in deep learning models for forecasting remains a significant challenge.

Purpose of the Study:

• To propose input significance and model validity tests for deep dendritic recurrent neural networks (DDRNNs) within an explainability framework.
• To develop a novel automated forecasting method for DDRNNs leveraging these proposed tests.
• To evaluate the forecasting performance of the new method against established techniques using benchmark datasets.

Main Methods:

• Development of statistical tests to assess the significance of inputs to DDRNNs.
• Implementation of model validity tests to ensure the reliability of DDRNN forecasts.
• Creation of an automated forecasting pipeline for DDRNNs incorporating the developed tests.
• Comparative analysis using M3 and M4 competition time series datasets.

Main Results:

• The proposed input significance and model validity tests demonstrate effectiveness in evaluating DDRNN components.
• The novel automated forecasting method shows competitive or superior performance compared to existing methods on M3 and M4 datasets.
• The developed tests contribute to the explainability of DDRNNs in time series forecasting.

Conclusions:

• The proposed tests enhance the interpretability and reliability of deep dendritic recurrent neural networks for forecasting.
• The new automated method offers a robust and data-driven approach to time series forecasting with DDRNNs.
• This work advances the field of automated forecasting by integrating explainability into deep learning models.