Prognostication of Sales by Auto Encoder and Long-Term Short Memory

Abstract

The intention of the paper is to improve a neural network methodology to accomplish enhanced predictions of the sales market. The data downloaded by Kaggle, data is surveyed for more than six months and the data was collected through prevalent markets for online and offline analysis with results of data visualization and prediction to illustrate sales forecasting. The traditional model like arima, RNN, and long short-term memory are not effective to provide sales forecasting with consideration of numerous constraints of the market and predict the sales incorrectly, because the RNN model suffers from vanishing gradient problems and LSTM are prone to overfitting. Therefore, these models are intensely prone to erroneous forecasts. The author suggests the “Long Term Short Memory (LSTM)” with three layers which are dropout layers, early stop layers, and simplifying layers to reduce overfitting. The result shows that the adapted “LSTM '' with the inclusion of three layers is an improved version as compared with traditional ''LSTM ``. The accuracy of the proposed model is 82%