This study develops a fashion sales trend prediction system to support retail decision-making by combining historical sales forecasting with contextual information from product captions. The dataset comprises 300 observations from 15 fashion products collected from January 2025 to August 2026. Text data were processed using Natural Language Processing and Term Frequency-Inverse Document Frequency to identify relevant keywords and contextual relevance to the target season. Sales forecasting employed a Seasonal Residual Multivariate Long Short-Term Memory model with a 12-month Lag-12 seasonal baseline and three numerical inputs: Terjual_Scaled, Month_Sin, and Month_Cos. An alpha value of 0.50 was selected using the validation set. On the testing set, the final model achieved an MAE of 10.358670, RMSE of 12.621822, and R² of 0.887345. The Flask-based system also provides monthly and seasonal growth, trend status, contextual insights, recursive forecasting, and visual product recommendations. Functional and logic testing confirmed that the prototype operated according to the designed scenarios.
Copyrights © 2026