CMStatistics 2016: Start Registration
View Submission - CMStatistics
B1378
Title: A comparative study on time series forecasting in telecommunications Authors:  Irene Castro-Conde - Optare Solutions S.L. (Spain) [presenting]
Marta Cousido Rocha - University of Vigo (Spain)
Javier Roca Pardinas - University of Vigo (Spain)
Antonio Vidal Vidal - (Spain)
Abstract: Telecommunications service providers have large quantities of historical data that could assist them to manage their business. In this sense, time series forecasting can assist decision making in marketing, planning or network management. In particular, we analyse real mobile daily data of a middle size operator corresponding to the number of people who left the company (churn) over the last four years. This is a problem of great interest since it is much cheaper to retain an existing customer than to acquire a new one and therefore retention strategies have to be developed in advance to avoid the churn. The time series under analysis presents multiple challenging characteristics such as multiple seasonalities and the influence of holidays and business cycle. The aim is to select the most suitable model in order to predict the next month occurrences. To this end we compare different methodologies which range from more classical approximations like ARIMA or Generalized Additive Models (GAM) to more general regression purpose methods like Recursive Partitioning Trees and Neural Networks. In order to do feature selection, parameter tuning and evaluate the results, three different error measures (MAE, RMSE and MASE) will be computed in a moving window validation scheme.