Customer Lifetime Value Prediction Using Long Short-Term Memory with RFM-T Features Based on E-Commerce Customer Data

Abstract

CLV prediction plays an important role in e-commerce by helping companies identify valuable customers and develop effective retention strategies. However, predicting CLV remains challenging due to the sequential nature of customer purchasing behavior. This study proposes a CLV prediction model using the LSTM algorithm with Recency, Frequency, Monetary, and Tenure (RFM-T) features extracted from the Online Retail II dataset. The CLV target is calculated as the accumulated monetary value generated during the three months following the historical observation period. The proposed approach consists of data preprocessing, monthly RFM-T feature engineering, feature normalization using min-max scaling, LSTM model training, denormalization, and performance evaluation using MAE and RMSE. The LSTM model incorporates stacked LSTM layers, batch normalization, dropout, and L2 regularization to improve learning stability and generalization. Experimental results indicate that the model was able to capture customer purchasing patterns based on sequential RFM-T features, with training and validation loss trends showing stable convergence. The proposed model achieved an MAE of 43.16 and an RMSE of 125.39 on the original monetary scale, reflecting the prediction performance obtained on the evaluated dataset. These findings suggest that the proposed LSTM model with RFM-T features provides an approach for CLV prediction in e-commerce.