Every roster in a contact centre starts as a guess about the future: how many contacts will arrive, when, and how long each will take. Workforce management teams call this the forecast, and almost every service level failure that is not caused by absenteeism traces back to it. Measuring how good that guess was sounds simple. In practice there are several ways to do it, and they tell different stories.

What a forecast is built from

A contact centre forecast usually predicts two things: volume (contacts offered) and average handle time, at the level of 15- or 30-minute intervals. Together they give workload, which drives staffing. The inputs typically include:

  • History: ideally a year or more of interval-level data, so that weekly patterns, month-end effects and annual seasonality are visible.
  • Patterns: day-of-week shapes (Mondays are heavy in many service processes), intraday curves, and pay-cycle or billing-cycle effects.
  • Known events: marketing campaigns, product launches, price changes, bill runs, festive sales such as the Diwali shopping season, tax filing deadlines, and public holidays that differ by state.
  • Business drivers: customer base growth, orders shipped, policies issued, anything with a reliable link to contacts.
  • Channel shifts: new self-service options, chatbot launches or IVR changes that move contacts away from or onto voice.

Methods range from simple (last year same week, adjusted for growth) to time-series models such as exponential smoothing and ARIMA, to regression on business drivers. A good forecaster also cleans history: a day when the IVR was down and calls could not get through is not a normal day, and including it untouched will distort future forecasts.

Measuring accuracy: a worked week

Here are five days of forecast and actual call volumes for an illustrative process:

DayForecastActualAbsolute % error
Mon1,2001,3007.7%
Tue1,5001,35011.1%
Wed1,1001,2109.1%
Thu9009808.2%
Sun30024025.0%
Total5,0005,080

MAPE, mean absolute percentage error, averages the daily percentage errors: (7.7 + 11.1 + 9.1 + 8.2 + 25.0) ÷ 5, about 12.2%. Notice how much the Sunday, a small day, drags it up. A 60-call miss on Sunday counts as heavily as a 150-call miss on Tuesday.

WAPE, weighted absolute percentage error, adds up the absolute errors and divides by total actual volume: (100 + 150 + 110 + 80 + 60) ÷ 5,080 = 500 ÷ 5,080, about 9.8%. Because it weights by volume, it reflects how many calls were mis-planned, which is closer to the operational impact. Many contact centres prefer WAPE, or report forecast accuracy as 100% minus WAPE, here about 90.2%.

Bias asks whether errors lean one way: (total forecast − total actual) ÷ total actual = (5,000 − 5,080) ÷ 5,080, about −1.6%. A small bias like this means over- and under-forecasts mostly cancelled out over the week. A persistent bias of, say, −8% month after month is more serious than random error, because it means the method is systematically wrong and the process will be chronically understaffed.

Note the sign convention: some teams compute actual minus forecast instead. Agree which one you use, or a “positive bias” will be read in two opposite ways.

The interval problem

A forecast can be perfect for the day and still wreck the service level. Imagine two half-hour intervals, both forecast at 100 calls. Actual arrivals are 150 and 50. The daily total is exactly right, an error of zero. But at interval level, absolute errors total 100 calls on 200 actual, a WAPE of 50%. The first interval was badly understaffed and the second overstaffed, and the queue in the first interval will have produced long waits and abandons that the idle agents in the second cannot fix.

So meaningful forecast accuracy reporting looks at several levels: monthly (for capacity and hiring), daily (for rosters and leave), and interval (for shift patterns and breaks). Interval accuracy will always look worse than daily, because small numbers are noisier. Targets should differ by level; a target that is reasonable for monthly volume may be unreachable at interval level.

Handle time is half the forecast

Teams often report volume accuracy and forget handle time. If volume is spot on but AHT runs 10% above forecast, workload is 10% higher, which has the same staffing effect as a 10% volume miss. AHT shifts after policy changes, system slowdowns, new joiner batches or a new product. Tracking workload accuracy, volume multiplied by handle time, captures both.

Using the error, not just reporting it

The point of measuring forecast accuracy is to improve the next forecast. Useful practices include recording a reason for every large miss (unannounced campaign, outage, weather, a viral social media post), separating misses the forecaster could have known about from genuine surprises, and feeding known events from the client’s marketing and product teams into the forecast earlier. A forecast that misses because the client did not share a campaign date is a communication problem, and no statistical model will solve it.

When accuracy cannot be pushed further, the remaining uncertainty should be handled through flexibility: multi-skilled agents, part-time or split shifts, voluntary overtime pools and intraday reshuffling. Good workforce planning accepts that the forecast will be wrong and designs the roster to absorb the error.

Written by the RP Strativa Team. Volumes and error figures are illustrative.

Leave a Reply

Trending