Prediction Modeling of Toy Sales Quantity Using Light Gradient Boosting Machine

Abstract

The main characteristic of the toy industry is its rapid change and uncertainty. Demand, influenced by certain trends, can change abruptly and suddenly disappear when the next viral product takes over the market. Constant product innovation, short life cycles and high cannibalization rates have the potential to incur higher relative costs compared to other industries in terms of inventory obsolescence, lost sales and reduced prices. Based on these problems, a study was conducted to predict toy sales using the LightGBM algorithm model in a time-series form with a sales dataset of 460 toy items classified into 14 categories within a time span of 1,353 days with a prediction period of 1, 3, and 6 months. This study produced 42 models based on product category and prediction period, with the best RMSE value of 0.0042 in the KARTU toy model, and 3 models for all categories based on the prediction period with the best RMSE value of 0.0380 in the 1 month prediction period.