Artificial Intelligence (AI) using Long Short-Term Memory (LSTM) for Sales Prediction in Campus Minimarkets

Abstrak

This study applies Artificial Intelligence (AI) using the Long Short-Term Memory (LSTM) algorithm to predict daily sales at the FIKOM-UMI Minimarket. Sales data from 2023 to 2024 involving 82 items were used and processed into a time series format. Five LSTM architectural scenarios were tested, including baseline, bigger model, lightweight, bidirectional LSTM, and single-layer medium, to identify the most effective model in capturing sales patterns. The data underwent preprocessing stages, including daily aggregation, reindexing to fill missing dates, and normalization using MinMaxScaler before being transformed into sequences with a 30-day time step. Model performance was evaluated using MSE, RMSE, MAPE, and accuracy metrics. The results show that the Bidirectional LSTM (Scenario 4) achieved the best performance, with the lowest MAPE of 19.43% and the highest accuracy of 80.57%. The model successfully generated stable predictions for 7-day and 30-day forecasting with a range of 153–155 units per day, indicating consistent sales patterns. Testing on the top 10 best-selling items showed significant performance variation, with GARUDA ROSTA BWNG 100 Gram achieving the highest accuracy (46.97%), while aoka rasa pandan showed the lowest performance (-76.05%). These findings demonstrate that the LSTM model can be effectively applied for sales prediction in campus minimarkets; however, a hybrid approach with product segmentation is recommended to optimize inventory management across product categories with varying levels of predictability