Accuracy Evaluation of Multi-Factor Forecasting Methods for Customer Service Workload Prediction Under Holiday and Promotional Fluctuations: Evidence from U.S. Service Industry Data

Abstract

Accurate prediction of customer service workload is essential for workforce scheduling and resource allocation in the U.S. service industry. This study conducts a comparative evaluation of four forecasting approaches—Seasonal Autoregressive Integrated Moving Average (SARIMA), Prophet, Multiple Linear Regression (MLR), and Random Forest—for customer service call volume prediction across normal operating periods and holiday/promotional fluctuations. Using publicly available call center datasets from the Technion Service Enterprise Engineering Laboratory and supplementary U.S. Bureau of Labor Statistics data, method performance is assessed using Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and Weighted Absolute Percentage Error (WAPE) at daily and intraday temporal granularities. The results indicate that all four methods exhibit notable accuracy degradation during holiday periods, with MAPE increases ranging from 67.3% to 118.3% relative to normal periods. Prophet demonstrates the smallest holiday-period degradation owing to its built-in holiday component modeling. At the daily level, SARIMA achieves the lowest overall MAPE of 7.82%, while Random Forest shows competitive performance at intraday granularity. Translating forecast errors into staffing implications via Erlang-C modeling reveals that improved holiday-period forecasting could reduce annual overstaffing cost waste by $88,000–$139,000 for a mid-sized 500-agent U.S. contact center.