Call Volume Forecasting

What gets forecast

Volume alone is not enough. A usable forecast has three parts for every queue and channel:

• Contact volume by interval (15 or 30 minutes).
• Average handle time by interval, since AHT often changes during the day and is higher for new hires.
• Arrival pattern, the shape across days and intervals.

These feed the staffing calculation, usually an Erlang C model or a simulation for multi-skill queues, which produces required agents per interval. Add shrinkage and you get the number to schedule. Planning that same chain months ahead to set hiring targets is capacity planning.

Common forecasting methods

• Moving averages and same-period-last-year. Simple and often good enough for stable queues. Take the last 4 to 8 comparable weeks, adjust for trend and known events.
• Exponential smoothing (including Holt-Winters). Weights recent data more heavily and handles trend and seasonality. This is the default in many workforce management platforms.
• Regression and driver-based models. Forecast calls from business drivers: orders shipped, active customers, bills sent, claims filed. Strong when the business is changing faster than history can show.
• Machine learning models. Useful with many queues and drivers, but they still need clean history and a planner who understands why the numbers look the way they do.

In practice, most good forecasts combine a statistical baseline with planner judgment about events the history does not contain: a new product, a price change, a system migration, a storm season.

How to measure forecast accuracy

The usual measure is mean absolute percentage error (MAPE): the average of |actual − forecast| ÷ actual across periods. Measure it at the level you staff, which means intervals, not just the month. A monthly forecast within 2% can still miss Monday mornings by 20%.

Common benchmarks are within about 5% at the daily level and within 10% to 15% at the interval level for queues with a few hundred calls per interval. Small queues will always look worse in percentage terms because random variation is larger. Track bias as well as error: a forecast that is always 6% low is easier to fix than one that is randomly off.

Forecasting for hiring and ramp-ups

Volume forecasts are also hiring forecasts. A center that expects a 25% increase in volume for open enrollment or the holiday season has to start recruiting 8 to 12 weeks earlier, because new agents need training and nesting before they count as full capacity. Forecast the expected call center attrition and training dropout too, or the hiring plan will be short.

For example, if the plan needs 120 productive agents by November 1, training attrition runs 15% and first-90-day attrition another 20%, the recruiting team needs to start roughly 175 hires across the waves, not 120.

Common forecasting mistakes

• Forecasting volume but not AHT. Handle time errors cause as much understaffing as volume errors.
• Ignoring abandoned and repeat calls. History built only on answered calls understates real demand when the queue was understaffed.
• Not removing one-off events from history. An outage week will distort next year's forecast unless it is flagged and cleaned.
• No feedback from marketing and operations. Campaigns, mailings and billing changes should reach the planner weeks ahead.
• Forecasting once a month and never adjusting. Reforecast the next two weeks as actuals come in, and let intraday staff react to the day.