Sales Demand Forecast based on Recurrent Neural Network∗
Zhaomin Zhu, Lin Lu, Ning Wu · 2022
In this paper, we compare the performance of RNN, LSTM and GRU, which are the most popular cyclic neural networks, in predicting the total sales of products in each store next month. A dropout layer is added to the model to reduce over-fitting. The results show that the root mean square errors between the predicted value and the actual value for all the three models are less than 0.47, while the value of GRU model is 0.39, which has the best performance. At the same time, LSTM and GRU prove that they are more robust to the gradient disappearance and gradient explosion problems of RNN. Compared with LSTM, RNN and GRU still have some over fitting problems.