Comparison of Deep Learning Methods for Product Sales Forecasting at The Catur Sasih Cooperative Four Seasons Hotel

Abstract

This study aims to assist Koperasi Catur Sasih Hotel Four Seasons in determining the most appropriate sales forecasting method based on the characteristics of the cooperative’s sales data, as fluctuating demand creates challenges in determining optimal inventory levels. The methods compared in this study are Long Short-Term Memory (LSTM) and Autoformer, using monthly sales data from January 2020 to December 2024 for three products, namely Yakult, Bavarois Roti Pizza/Sisir, and Marlboro Lights 20. The results indicate that the Autoformer method provides more accurate sales predictions than the LSTM method in forecasting product sales at Koperasi Catur Sasih, as evidenced by the lowest error values according to the RMSE, MAE, and MAPE metrics. For the Yakult product, the Autoformer method achieved an RMSE of 82.4660, an MAE of 72.2140, and a MAPE of 9.65%. For the Bavarois Roti Pizza/Sisir product, the Autoformer method produced an RMSE of 18.2666, an MAE of 15.4194, and a MAPE of 11.37%. Furthermore, for the Marlboro Lights 20 product, the Autoformer method resulted in an RMSE of 44.3490, an MAE of 37.0785, and a MAPE of 12.36%. In addition, noise reduction testing by removing the first three months of 2020 as an anomalous data period resulted in a decrease in error values for both methods. Nevertheless, Autoformer consistently outperformed LSTM in sales forecasting, both before and after the noise removal process.