Customer lifetime value prediction for an online marketplace for local services using machine learning

In this age of direct and digital marketing, customer-oriented metrics, especially customer lifetime value (LTV), have become one of the most important business metrics for customer relationship management (CRM) and marketing strategies in general. The customer LTV, a quantification of the profit or revenue that a customer will bring to the firm over their entire relationship period, differs wildly for different customers, which makes predictions over the entire pool of the firm's customers quite challenging. The rapidly expanding field of machine learning is uniquely suitable for solving this problem. In this study, we develop a model for predicting customer LTV for a local services marketplace, namely Armut AŞ. The predicted LTV is then used for the optimization of the company's digital ads. The model was developed through experimenting with several machine learning algorithms including linear regression, decision tree, extreme gradient boosting (XGBoost), and artificial neural networks (ANNs), as well as experimenting with a combination of encoding techniques for the categorical features in the data set. The models were evaluated using 5-fold time-series cross-validation and yielded comparable results with normalized root mean square error (NRMSE) ranging from 1.740 to 1.776. The chosen model was XGBoost, and it was further optimized by segmenting the data based on its business model as well as expanding the training set time window to 12 months yielding the final model with an NRMSE of 1.725. This model was deployed and is currently operating at Armut AŞ.