Comparative study of AutoML and machine learning in retail forecasting

Accurate retail sales forecasting is crucial for effective inventory management, operational efficiency, and ensuring customer satisfaction. While traditional machine learning models such as SARIMAX, Prophet, and Gated Recurrent Units (GRU) require extensive manual tuning and domain expertise, the recent emergence of Automated Machine Learning (AutoML) frameworks, including H2O AutoML, FLAML, and AutoGluon, promises to simplify and accelerate the forecasting pipeline. This study systematically compares these Traditional models and leading AutoML frameworks using an authentic retail sales dataset consisting of approximately 27,300 transaction records from an Indonesian traditional store. Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). Our results show that SARIMAX consistently delivers the most accurate forecasts for structured, seasonal retail data, outperforming both deep learning and AutoML approaches in this context. Among the AutoML frameworks, H2O AutoML offers the best trade-off between accuracy and the ability to capture sales volatility, while AutoGluon provides stable predictions, and FLAML shows higher error under volatile conditions. This study provides practical insights for both researchers and retail practitioners, highlighting that while traditional models remain highly effective for non-ideal, real-world data with strong seasonality, modern AutoML frameworks offer substantial advantages in automation and ease of use, making advanced forecasting more accessible in diverse retail environments.