Forecasting of Sales Based on Long Short Term Memory Algorithm with Hyperparameter
Ali Khumaidi, Ika Ayu Nirmala, Herwanto Herwanto · 2022
Increasingly competitive business competition requires business people to re-design their business strategies, one of which is by applying forecasting methods. Several forecasting techniques have been used in sales prediction research with fairly good accuracy in a short period, the research will focus on optimizing the hyperparameter LSTM algorithm to improve the performance of the model formed over the next 60 days. The method used in this study is the Cross Industry Standard Process for Data Mining (CRISP-DM), in the preprocessing the data cleaning, labeling, summary, and data transformation. The data understanding stage applies the Exploratory Data Analysis method. The development of the LSTM model uses several parameters, namely data partition, number of hidden layers, dropout scenarios to prevent overfitting, number of neurons, epoch describing the number of training iterations, batch size is the amount of training data that must be considered in each process of updating the weights. The experimental results of the best LSTM model after experimenting with different parameters are hyperparameter batch size 30, epoch 150, 3 hidden layers and 3 dropouts, resulting in RMSE training of 0.0855 and RMSE testing of 0.0846.