Forecasting System Analysis using Gated Recurrent Unit Neural Network

Abstract

Sales competition is tightening, and people's wants and wishes are becoming more complex. Retail organizations must plan strategically when buying items. Retail X Company started in 2004. Due to lack of planning, the warehouse's stock of items runs out quickly. Purchasing division in this company usually relies on instinct when purchasing items for stock. It always becomes a problem by using this method because that make inaccuracy in stock planning. For better planning, forecasting the number of items sales is needed so that planning for purchasing inventory can be more accurate to reduce the risk of unsold items due to excess purchases. Forecasting system is a tool or methodology used to predict future items sales based on historical data. It is faster and more efficient for processing sales forecasting. In this research the forecasting system used the Gated Recurrent Unit (GRU) model to anticipate product sales for more accurate inventory planning. The model was trained and validated using 2017-2021 time series data. The best model has batch size 64 and 64 hidden neurons, with MSE train scores of 138.8019 and validation scores of 136.3658. Using January 2022 and February 2022 sales data as actual data, MAPE evaluations were 46% and 60%. With this finding, Retail Company X can use GRU Neural Network to get reasonable forecasting for sales forecasting.