Layered Ensemble Architecture for Time Series Forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Time series forecasting (TSF) is crucial in science, engineering, and finance.
• Accurate forecasting requires selecting an optimal number of past values (lag).
• Existing TSF methods often struggle with unknown underlying phenomena and limited historical data.

Purpose of the Study:

• To propose a novel Layered Ensemble Architecture (LEA) for enhanced time series forecasting.
• To address the challenge of optimal lag selection in TSF.
• To improve forecasting accuracy by combining network diversity and accuracy.

Main Methods:

• Developed a two-layer ensemble architecture (LEA).
• The first layer uses an ensemble of multilayer perceptron (MLP) networks to determine the appropriate lag.
• The second layer uses another MLP ensemble, leveraging the determined lag for forecasting.

Main Results:

• The proposed LEA considers both accuracy and diversity in ensemble construction.
• Networks within the ensemble are trained on different datasets to ensure diversity.
• LEA demonstrated superior forecasting accuracy compared to existing ensemble and non-ensemble methods in extensive tests.

Conclusions:

• The Layered Ensemble Architecture (LEA) offers a significant improvement in time series forecasting accuracy.
• LEA effectively balances network diversity and accuracy for robust TSF.
• The method shows strong performance on benchmark datasets and competition data.