Design of an analysis application for predict demand forecasting in inventory system

Abstract

Trading companies often face challenges in managing their inventory systems and finding it hard to match inventory levels with the increasing demand from consumers. This paper tries to create a system that can forecast how much of a product customers might want and also check how well different deep learning methods work. In this paper, the author uses both experimental and applied methods to check how well each deep learning algorithm works.The algorithms tested are CNN-LSTM, LSTM, GRU, BiLSTM, and Transformer. The author used data from Kaggle that includes 76,000 sales records from 2022 up to 2024, covering different product categories such as clothing, furniture, electronics, toys, and groceries. The data was then cleaned and split into groups for training and testing. The author found the results by testing various deep learning algorithms. The top performing algorithm is CNN-LSTM, which has MAE of 28.2136, RMSE of 36.6037, and R² of 0.3106. Even though the R2 value shows the model doesn't have strong predictive power yet, it is still the best performing model out of all the ones studied in this research. By comparing the system's performance to standard operating levels, it was found that the system can adjust to changes in demand.This suggests it could help trading companies better manage their stock and offer more understanding of how deep learning models work.