Forecasting Seasonal Sales in Supermarket Chains Using FFT-GA and LightGBM Models

Jui-Chung Hung

Abstract:

This paper presents a two-stage prediction model designed to effectively address the fluctuations in sales volume experienced by supermarket chains due to seasonal patterns, sales cycles, holidays, and irregular variations. In the initial phase, the model combines raw sales data with Fast Fourier Transform (FFT) to uncover seasonal characteristics and then employs a Genetic Algorithm (GA) with its binary encoding capabilities for feature selection.
In the subsequent stage, the model leverages the Light Gradient Boosting Machine (LightGBM) as the prediction model, using weak learning strategies to manage irregular variations in sales data. Empirical simulations using three publicly available datasets of supermarket chains from Kaggle validate the effectiveness of our proposed method. This approach holds significant practical implications for understanding and predicting fluctuations in supermarket sales data and offers a new direction for future research in other domains.