Predicting the number of customer transactions using stacked LSTM recurrent neural networks

Time series forecasting is used to predict future values on a data sequence using time-based data. Time series of transactions or count data are difficult to predict because of their complex nonlinear patterns. The purpose of this study is to provide a solution for predicting the number of upcoming transactions in order to predict the anomaly of the system, with the purpose of checking for why the number of transactions in the next day to be lower than expected. This can be a kind of customer churn prediction and can help system administrators identify and prevent potential losses and can improve existing infrastructure. Required data were collected from 780 company, which contains 353 rows of transactions in 2018 and has two variables, date and number of transactions; the number of transactions has been set as a target to predict. The autoregressive integrated moving average (ARIMA) model cannot deal with nonlinear relationships, while the neural network model, conversely, is capable of processing nonlinear patterns. In this study, a deep learning approach is presented. The approach is a stacked long short-term memory (LSTM) structure, a model derived from recurrent neural networks. To find the most accurate prediction model, the performance measures of various recurrent neural network, PROPHET, and ARIMA models are compared using the same dataset. Experimental results show that for predicting the number of transactions, the stacked LSTM model is better than other approaches.