Daily Sales Prediction for TikTok Shop Using a Hybrid Stacking Regressor: A Case Study of OnefreshID

Abstract

Predicting daily sales on TikTok Shop remains challenging because sales patterns are influenced by userinteraction dynamics, temporal characteristics, and non-linear demand fluctuations. In social commerce, interaction metrics such as likes, comments, shares, and views can serve as early indicators of audience interest, while temporal factors such as day of week and date period provide contextual information on purchasing behavior. This study analyzes the predictive contribution of user-interaction variables and temporal factors to daily TikTok Shop sales for OnefreshID and develops a Hybrid Stacking Regressor that combines Long Short-Term Memory (LSTM), one-dimensional Convolutional Neural Network (CNN 1D), and Random Forest as base learners, whose outputs are combined by a meta-learner. The model is evaluated using a time-based data split with Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-squared metrics, while Permutation Importance and SHapley Additive exPlanations (SHAP) are used to interpret feature contributions. Results show that the Hybrid Stacking Regressor outperforms all single models, with MAE of 23.33, RMSE of 30.10, and R-squared of 0.57. Views and likes are the most dominant predictive features, while temporal factors act as supporting context. These findings indicate that OnefreshID’s daily sales are explained more by content performance and audience response than by calendar effects alone.