Multi-Horizon LSTM Forecasting of Digital Product Transactions in Retail Counter

Abstrak

This study proposes a multi-horizon forecasting approach using a stacked Long Short-Term Memory (LSTM) neural network to predict digital retail transaction volumes in a micro-scale counter environment. Historical transaction data were obtained from MyFastbiller Telegram logs and transformed into a 739-day daily time series containing 27,540 valid transactions. Temporal feature engineering, normalization, and 60-day sequence generation were applied prior to model training. The LSTM architecture was evaluated across three operational forecasting horizons: H-1, H-7, and H-30.

Results show consistent prediction accuracy with MAPE values of 15.24%, 15.14%, and 16.35% for H-1, H-7, and H-30 respectively, all below the 20% operational threshold. The H-7 horizon achieved the best performance due to alignment with weekly seasonality patterns in transaction behavior. Visualization results indicate that predicted transaction series closely follow actual temporal dynamics across all horizons.

The trained model was integrated into a web-based decision-support dashboard to provide multi-horizon transaction forecasts for operational planning. The findings demonstrate that multi-horizon LSTM forecasting effectively captures seasonal and intermittent characteristics of digital retail transactions and provides practical value for data-driven balance management in micro-scale retail counters.