Enhanching Prophet Time Series Forecasting on Sparse Data via Hyperparameter Optimizattion: A Case Study in Retail

In today’s competitive business landscape, accurate sales forecasting is crucial for retailers to optimize inventory, prevent overstock, and support strategic decision-making. However, many small to medium enterprises operate with sparse and irregular sales data, making conventional forecasting methods less effective. This study aims to evaluate the performance of the Prophet time series model in such non-ideal conditions and to investigate how hyperparameter tuning affects its forecasting accuracy. The research adopts the Prophet algorithm, an additive time series forecasting model developed by Facebook, which incorporates trend, seasonality, and holiday components. The model was implemented in two configurations: one using default parameters, and another with manually tuned hyperparameters, including changepoint prior scale (CP), seasonality prior scale (SP), and seasonality mode. A total of 32 experiments were conducted using historical transaction data from PT Eko Hejo. Results show that the default Prophet model achieved a MAPE of 9.50%, while the best-performing configuration (CP = 0.5, SP = 0.01, additive mode) reduced the MAPE to 6.80%. This indicates that hyperparameter tuning significantly improves forecast accuracy, even in sparse data environments. The study contributes both practically and scientifically by demonstrating that Prophet, when properly configured, is a robust and adaptable tool for business forecasting with limited data. It also highlights the value of manual tuning in enhancing model responsiveness and generalization, offering insights for further research in model comparison, automated optimization, and hybrid forecasting approaches.