Raw Paper Material Stock Forecasting with Long Short-Term Memory

Abstract

The manufacturing business is one of the businesses in Indonesia that continues to show its development from year to year. Like a manufacturing business in general, one of the important efforts made in the printing business is the supply of raw paper materials to produce finished goods. The purpose of this research is making a forecasting of the raw paper material for printing company on 7 different types of 269 historical data with weekly intervals from January 2015 to February 2020 before the Covid19 pandemic season. Forecasting is done using the Long Short Term Memory method with Python language. The model architecture for training and testing is carried out using vanilla LSTM with single input, hidden and output layer with the configuration of 64 neurons in the hidden layer, 150 epoch, 12 batch size and Adam Optimizer (lr = 0.0001) which was repeated 10 times for best result. The test results show the best window size length in the model for each paper raw material differently from 4 to 16. All models was successfully forecasting the test data with an average MAPE of the overall forecast of 21.48%.