Time series forecast of call arrivals using machine learning methods

Abstract

This study focuses on enhancing workforce management in the Citizen Service Request (CSR) Call Center dataset of the government of Cincinnati, Ohio, by improving the accuracy of call arrival forecasts. Recognizing the pivotal role of precise call arrival predictions in optimizing call center operations, this the study conducts experiments by utilizing a range of forecasting models, including statistical, machine learning, and neural network approaches. Feature engineering was proposed to broaden the scope of features for forecasting. The top-performing models are evaluated based on key metrics such as Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and R-Squared (R²) forecasting performance. The experimental results highlighted the comparative performance of various models, such as SARIMAX, Light Gradient Boosting Machine (Light GBM), Gradient Boosting Regressor (GBR), eXtreme Gradient Boosting (XGBoost), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). Among these, Support Vector Regression (SVR) leads in accuracy with an MAE of 25.13, an MAPE of 6.15%, an RMSE of 34.46, and an R² of 90.56%. The features of abandon rate, answer speed, service level calls, and the 1st and 5th lags, were identified as the most importance feature in this research. These findings provide valuable insights for the improvement of workforce management strategies in call center operations, emphasizing the effectiveness of machine learning algorithms in achieving more accurate call arrival forecasts.

-