A new automatic forecasting method based on a new input significancy test of a single multiplicative neuron model...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Single multiplicative neuron model artificial neural networks lack established model adequacy and input significance tests.
• Existing forecasting methods for these networks are not systematically based on hypothesis testing.

Purpose of the Study:

• To introduce novel hypothesis tests for model adequacy and input significance in single multiplicative neuron models.
• To develop a systematic, hypothesis-test-driven approach for artificial neural network forecasting.
• To propose an automatic forecasting method integrating these tests with optimization algorithms.

Main Methods:

• Development and simulation-based investigation of new model adequacy and input significance tests.
• Application of proposed tests to real-world forecasting examples.
• Integration of tests with particle swarm optimization for an automatic forecasting system.

Main Results:

• Simulation studies demonstrate excellent performance of the proposed test procedures.
• The developed automatic forecasting method ranked as the top pure machine learning approach in the M4 competition.
• The proposed method surpassed benchmark models like MLP, RNN, and ETS in accuracy.

Conclusions:

• The proposed hypothesis tests provide a robust framework for specifying and validating single multiplicative neuron models.
• The automatic forecasting method offers a highly accurate and competitive solution for time series forecasting.
• This work bridges a gap in the systematic application of hypothesis testing for neural network-based forecasting.