Related Experiment Videos
Meta-LLSTM: meta-learning enhanced learnable LSTM for retail sales forecasting
B S Suresh1, M Suresh2, Dae-Ki Kang3
1Department of Management Studies, St. Peter's Institute of Higher Education and Research, Chennai, India.
Scientific Reports
|May 25, 2026
Summary
The Meta-Learning Enhanced Learnable Long Short-Term Memory network (Meta-LLSTM) improves retail sales forecasting accuracy. This novel approach enhances business strategies by effectively handling linear and nonlinear data transformations for better revenue.
Area of Science:
- * Data Science and Machine Learning
- * Business Analytics and Operations Research
Background:
- * Accurate retail sales forecasting is vital for inventory management, demand prediction, and strategic business planning.
- * Conventional forecasting models often fail to capture both linear and nonlinear dynamics in time series data, limiting their adaptability.
- * Integrating precise forecasting with effective business strategies is key to enhancing company revenue.
Purpose of the Study:
- * To introduce the Meta-Learning Enhanced Learnable Long Short-Term Memory network (Meta-LLSTM) for superior retail sales forecasting and generalization.
- * To address limitations in existing models by incorporating meta-learning for rapid adaptation and a novel activation function for handling complex data transformations.
- * To enhance forecasting accuracy and business strategy development through integrated analytical modules.
Main Methods:
- * Development of the Meta-LLSTM model, utilizing meta-learning for adaptability and the Multiple-Parameter Exponential Linear Unit (MPELU) for handling linear/nonlinear transformations.
- * Integration of Recency, Frequency, Monetary, and Diversity (RFMD) analysis for customer sales data.
- * Implementation of K-means clustering for customer segmentation and Adaptive Inventory Correction (AIC) for inventory data management.
Main Results:
- * The Meta-LLSTM model demonstrated superior performance, achieving a Root Mean Square Error (RMSE) of 1.003.
- * Achieved a 16.97% reduction in RMSE compared to the Recurrent Neural Network (RNN), a leading state-of-the-art method.
- * Outperformed other advanced models including Auto Encoder (AE), Convolutional Neural Network (CNN), and Gated Recurrent Unit (GRU) in forecasting accuracy.
Conclusions:
- * The Meta-LLSTM model offers a significant advancement in retail sales forecasting, outperforming existing state-of-the-art methods.
- * The integration of meta-learning and advanced activation functions enables robust handling of complex time series data.
- * The proposed model provides a foundation for more effective business strategies and improved revenue through enhanced forecasting accuracy.