Creative Approaches Towards Development of Computing and Multidisciplinary IT Solutions for Society

15An Efficiency Improvement of the N-Beats Model for Sale Forecast Problem

Huy Nguyen Dinh^1,2, Trong Hop Do^1,2 and Thanh Binh Nguyen^1,2*

¹University of Information Technology, Ho Chi Minh City, Vietnam

²Vietnam National University, Ho Chi Minh City, Vietnam

Abstract

Deep-learning models have shown remarkable success in time-series forecasting tasks in recent years. One such state-of-the-art model is N-Beats [1, 2] (A Neural basis expansion analysis for interpretable time series forecasting), which employs an interpretable architecture to capture the underlying patterns in time series data. However, despite its effectiveness, N-Beats can benefit from further customization to improve its predictive performance.

This study proposes a novel improvement method for N-Beats to achieve improved predictive results. Our approach focuses on one aspect: the model architecture. We refine the architecture of N-Beats by changing the input value calculation formula for each Stack and Block of the N-Beats network structure.

To evaluate the effectiveness of the proposed improvement, we experimented with three N-Beats models: N-Beats-I, N-Beats-G, and N-Beats-I-G, and compared them with the corresponding models with an improved architecture. The experimental process uses the sales data of a pharmaceutical distributor in Vietnam to predict the future sales volumes in future of that company. The experimental results show a better performance with a prediction accuracy of 11.1% than the original ...