Time Series Forecasting of Active Customers Using Sequence Models: A Comparative Evaluation

Active customer forecasting is crucial for software-as-a-service (SaaS) companies to plan resources and understand customer dynamics. This study benchmarks sequence modeling approaches to predict active customer accounts using a real-world dataset. It evaluates hybrid recurrent and convolutional neural networks against Facebook Prophet models, focusing on both multivariate and univariate time series analysis. Advanced models like Long Short-Term Memory (LSTM) networks and 1D Convolutional Neural Networks (Conv1D) are utilized. A comprehensive preprocessing pipeline and Bayesian hyperparameter optimization ensure robust models. The multivariate LSTM-Conv1D model shows superior performance, with a test mean absolute error of 16.49 and an R-squared of 0.84, outperforming Prophet models and univariate LSTM-Conv1D. The hybrid deep learning architecture excels by incorporating multiple related time series and modeling their complex interactions. The study finds up to a 96.62% improvement in accuracy over the Prophet model, highlighting deep learning's capability to capture intricate customer dynamics. This research provides practical guidelines for effective nonlinear time series modeling in customer forecasting for SaaS companies, enhancing data-driven decision-making.

Time Series Forecasting of Active Customers Using Sequence Models: A Comparative Evaluation Keywords:

Machine Learning , Neural Networks , Time Series Analysis , Bayesian Optimization , Customer Relationship Management , Customer Experience