Prediction of Customer Purchases Using LSTM Deep Neural Network

In the realm of business, accurately predicting customer purchases holds significant importance, as it enables the adept tailoring of marketing strategies, enhances promotional campaign effectiveness, optimizes inventory management, and reduces food loss and waste, ultimately leading to sales growth and profit maximization. The paper aimed to develop a deep learning model using the LSTM (long short-term memory) neural network to predict customer purchases in the next week, including transaction count and overall transaction gross value, based on their historical transaction data. In the research, to achieve the best predictions, various sizes of the training dataset, different model architectures, and its hyperparameters, as well as the hyperparameters of the learning algorithm, were tested. Prediction performance was assessed by RMSE and R2. As a result, the optimal predictive model trained using 10% of customers’ transaction sequences was characterized by very high performance at the levels of 1.4 (for transaction count) and 26.2 (for transaction gross value) in RMSE and at the levels of 99.9233% (for transaction count) and 99.4425% (for transaction gross value) in R2. The obtained model can be implemented before validating its performance on another instance of production data.