Forecasting Calling Activity Based on ARIMA Model

Call frequency time series are extracted from Call Data Record and widely used to explore human activity patterns as well as predict calling activity. Accurate call frequency forecasting is a fascinating and challenging work. In this paper, we propose a recursive approach based on ARIMA model to forecast short term call frequency. This approach contain three steps. First two steps are modeling process and the third one is forecasting. The main work of step1 is data preprocessing and determining d (differential order) in ARIMA model. Further, a series of potential p (auto-regressive order) auto-regressive parameters values and q (moving average order) parameters values can be determined by examining the ACF (Auto Correlation Function) diagram and PACF (Partial Auto correlation Function) diagram. Then, the minimum BIC (Bayesian Information Criterion) rule is used to determine a p parameter value and a p parameter value. In the third step, the next day call frequency is predicted and all parameters are updated in a automatically recursive manner. We apply this approach on 10 call frequency time series and analyze forecasting error. The average MAPE (Mean Absolute Percentage Error) is 4.792%, which is much lower than other methods based on ARIMA model. The results show that our approach is efficient and can be used to in practical work.