Forecasting call center arrivals using temporal memory networks and gradient boosting algorithm

For any call center facility, the number of call arrivals represents a key component between customer satisfaction and budget constraints. Hence, the ability to accurately forecast the number of calls for a particular period of time is an effective measure in planning customer service and reducing any resource waste. In this research, presented as a comparative study in a tutorial style, we present call arrival forecasting comparison between classical time series methods such as Holt-Winters, exponential smoothing, and SARIMA with machine/deep learning techniques like Recurrent Neural Networks and gradient boosting approach. To test the models, we use real-life call logs from a national US insurance company collected for ten months. The series exhibits an inhomogeneous Poisson process and dependencies between aggregated periods. Call center managers commonly require forecast results on short and long-term periods to determine the required headcount based on expected service quality. In our case, we used half-hourly and daily aggregations for short and long-term forecasts. Because the data used is less than a year long, we provided enough seasonal periods in the short-term period. In this case, both deep learning models reported minimum error followed by the boosting approach. This is not the case for long-term periods, where the provided series is less than a year. The boosting approach reports better error results than any of the models used, even deep learning models which report the worst errors from the model's list. Those results indicate that on limited seasonality periods, deep learning models are incapable of generalizing different volatility fluctuations compared to boosting and classical approaches. Considering boosting errors on short-term periods are comparable to deep learning models, we suggest the use of boosting as a benchmark approach to forecast call arrivals for the inhomogeneous Poisson process, specifically on limited seasonal periods.

Automated summary

For call center facilities, accurately forecasting the number of call arrivals is crucial for customer satisfaction and budget management. This study compares classical time series methods with machine/deep learning techniques to forecast call arrival, using real-life call logs from a national US insurance company. The results show that deep learning models perform well in short-term periods with enough seasonal data, but boosting approach outperforms all models, including deep learning, in long-term periods. These findings highlight the importance of considering limited seasonality and the use of benchmark approaches for call arrival forecasting.