Telecommunications Product Revenue Time-Series Forecasting Using Target Variable Preprocessing Methods

Authors

Philippe Anthony C. Bautista, Ateneo de Manila UniversityFollow
Christian Paul O. Chan Shio, Ateneo de Manila UniversityFollow
Patricia Angela R. Abu, Ateneo de Manila UniversityFollow

Document Type

Conference Proceeding

Publication Date

3-5-2025

Abstract

Accurate revenue forecasting is critical for decision-making support of telecommunications companies (telcos). This study explored the use of machine learning for time-series revenue forecasting of a telco product. While existing research explores machine learning for time-series forecasting primarily for stocks price prediction and different use cases in other industries, this study focused on telco revenue and the impact of target variable preprocessing on forecasting accuracy. Two datasets with different business rules for the same attribute were used, with two preprocessing techniques for converting monthly revenue data to daily applied to each dataset: even distribution and a weighted distribution based on daily subscriber count. Recurrent neural networks (RNNs), specifically long short-term memory (LSTM) and gated recurrent unit (GRU), were employed for revenue prediction. Various factors were used to produce additional model variations, specifically seed selection, dataset, preprocessing technique, and input window size. Mean Absolute Percentage Error (MAPE) was the key metric for comparison of model performance. The results showed that weighted preprocessing produced the most accurate model with an MAPE of 3.08% despite its reliance on a specific variable combination. This study concludes that target variable preprocessing impacts model outputs, with weighted distribution offering the highest accuracy for telco product revenue forecasting using RNNs.

Recommended Citation

Philippe Anthony C. Bautista, Christian Paul O. Chan Shio, and Patricia Angela R. Abu. 2025. Telecommunications Product Revenue Time-Series Forecasting Using Target Variable Preprocessing Methods. In Proceedings of the 2024 6th Asia Conference on Machine Learning and Computing (ACMLC '24). Association for Computing Machinery, New York, NY, USA, 53–61. https://doi.org/10.1145/3690771.3690784