Elder Research partnered with the call center leadership of a leading food service chain to optimize staffing levels by understanding and forecasting call center volume and staffing needs. The call center’s unique approach to handling outbound calls meant their off-the-shelf solution did not provide adequate forecasts of call volume or staffing needs.
The client’s call center is an internal service desk for franchise owners to ask questions about daily operations of their franchise location. While calls can encompass any number of topics, this project was focused on calls related to information technology and related processes. Historical call center trends were analyzed to deploy a forecast of call center demand and staffing needs to inform call center operations.
During the data discovery phase, we reviewed existing process documentation and had discussions with call center analysts and business process owners to understand the available data sources, the current staff schedules, and the expectations for forecasting. Quantitative data sources included:
- Call logs – information about each call, including the type, duration, and the time spent waiting in queue
- Agent data – information about the agents handling calls, including the time spent on call, waiting to take calls, or doing work related to calls
- Incident data – information about ticket-level metrics such as first level resolution rate and time to resolution
To ensure we were working with representative data, we worked with the client to focus forecasting efforts on the most relevant data, account for outliers, and resolve instances in which data was missing from the base table. In our exploratory analysis, we sought to understand call volume trends, call handling time, and service level metrics such as speed of answer and percent of calls queued.
Using historical call volume data, we built a model of future inbound and outbound call volume at 2-week and 1-month horizons, and cross-validated the model across a range of training and forecasting dates using the mean absolute error as a measure of forecast quality. We considered time series modeling techniques such as TBATS, linear modeling, seasonal naïve, ARIMA, and Prophet. During model selection, seasonal naïve models performed well for both inbound and outbound call volumes due to the consistent weekly and daily pattern of call arrivals.
The staffing forecast built upon the call volume forecast using the Erlang C formula, which calculates the probability of a call being sent to the queue given the call workload and number of agents staffed at a given time. Service level targets along with the queueing probability were used to determine the optimal number of agents to staff throughout the day.
The forecasts were visualized in Tableau along with descriptive statistics regarding the historical call volume, staffing levels, and service level metrics. The staffing estimates were also applied to the historical call volume data and compared to the actual number of agents staffed. The example below shows the difference between the actual versus forecasted number of agents for a service level of 80% of calls answered in 60 seconds. A negative number indicates understaffing.
The resulting Tableau dashboards provide valuable insights into call volume, handle time, and service level metrics in addition to displaying the forecasts of call volume and staff needed. This information can impact future hiring and scheduling of call center agents and allow for monitoring of call volume and service level, ensuring service level goals are reached as call center volume grows.
Our analysis also provided valuable insight into the current business process and data pipeline, highlighting additional opportunities to improve the operations and flow of data.