Why Forecast Accuracy Matters
Forecast accuracy plays a central role in contact center planning. Arrival forecasts determine staffing levels, schedules, and day-to-day operational decisions. When forecasts are too low, queues grow, wait times rise, service levels fall, and agents come under pressure. When forecasts are too high, too many hours are scheduled, productivity drops, and costs increase. Good forecasting helps balance customer experience and efficiency.
WAPE as a Metric for Overall Accuracy
In terms of overall forecast accuracy, a good and easily applicable measure is WAPE (Weighted Absolute Percentage Error). This is calculated as the sum of all absolute forecast errors, divided by total actual volume. This is a single figure, showing overall forecast accuracy as a percentage. It is weighted, because it is divided by total actual volume. In contact centers, it is often more important for forecast accuracy to be good on busy days or busy skills, because errors on these days or skills are more important than errors on quiet days or quiet skills.
WAPE is a powerful measure for overall forecast accuracy because it is stable, easy to understand, and does not allow small errors in quiet intervals to dominate.
Although WAPE is useful for measuring overall forecast accuracy, it does not show how errors are distributed across the operation. In larger operations with many departments or work types, it is also helpful to evaluate forecast accuracy at the skill level. This is because a poor forecast for one skill can cause issues even if the overall center forecast is accurate. This matters in practice because forecasting is often not only about getting the total right. It is also about knowing where the forecast is consistently reliable, and where it is not.
The Challenge of Comparing Skills Fairly
When trying to compare forecast performance fairly between skills of very different sizes, WAPE becomes harder to interpret because it is strongly influenced by volume. In low-volume skills, even a small absolute error can be a large percentage when divided by a small denominator. In high-volume skills, large absolute errors can be small percentages.
As an example, suppose we have a high-volume skill where the actual arrivals are 1,000, and the forecast is 950. This gives an absolute error of 50 and a WAPE of 5%. Suppose we have a low-volume skill where the actual arrivals are 20, and the forecast is 25. This gives an absolute error of 5, but a WAPE of 25%. The low-volume skill looks much worse than the high-volume one, even though the absolute error is much smaller.
So when comparing skills, the real question becomes: are we measuring forecasting quality, or are we partly just measuring scale?
This is an important distinction. A metric can be very useful for one purpose, but less useful for another. WAPE works very well as an overall operational accuracy metric. But if the goal is to compare and rank forecast quality across skills of very different sizes, we may want an additional lens.
Why Volume Distorts Comparisons
One reason this is challenging is that contact center arrivals do not behave the same way at different volume levels. A skill with 20 contacts in an interval naturally behaves differently from a skill with 200 contacts. Larger-volume series tend to show larger raw swings simply because they are larger. That means a raw error of 20 contacts does not mean the same thing in every skill. This is well known in forecasting and statistics. If we want to compare forecast performance across series fairly, we often need some form of scale normalisation.
A Different Lens: Square-Root Transformed Error
One possible way to handle this is to evaluate forecast errors on a transformed scale rather than the original arrival scale. The idea is simple: instead of asking only âhow many contacts was the forecast off by?â, we ask âhow large was the miss relative to the natural scale of that skill?â One useful transformation for count-type data, such as arrivals, is the square-root transformation.
If we apply a square-root transformation to both actual arrivals and forecasts, we can compare them on a different scale:
where At is the actual number of arrivals in interval t, and Ft is the forecast.
The purpose is not to replace WAPE. The purpose is to create a more balanced way to compare skill-level forecast performance when volumes differ strongly. The intuition is that the square-root scale compresses larger volumes more than smaller ones. That reduces the extent to which size alone drives the comparison.
Why This Transformation Makes Sense
There is also a statistical reason this approach is useful. When series differ substantially in scale, a transformation can make low- and high-volume count-like series more comparable by reducing the extent to which variance and raw size dominate the error comparison. The appropriate transformation depends on the underlying distribution of the data. For example, for count processes that follow a Poisson distribution, such as arrivals, the square-root transformation is commonly used, as it helps stabilise the variance. More generally, in forecasting practice, transformations are applied when variability changes with the level of the series, and Hyndman (2014) notes that such transformations, including the square root, are standard tools for variance stabilisation.
A Worked Example
This transformed error can be easier to compare across skills, because it is less directly driven by size alone. A miss in a small skill and a miss in a large skill can be judged more on how meaningful they are relative to the natural variation of that skill, rather than on raw volume alone.
For example:
- Skill A: forecast 16, actual 25
- Skill B: forecast 196, actual 225
On the raw scale, Skill A misses by 9 and Skill B by 29, so Skill B looks much worse. But on the square-root scale:
So on the transformed scale, the misses are equally large.Â
That is the key idea: these two forecast misses are not equally large in raw contacts, but they are equally large relative to the scale of the skill.
Visualising the Difference
The charts below help illustrate the issue. In the first two charts, we compare a low-volume series and a high-volume series over time. Even when both forecasts are of similar quality in relative terms, the high-volume series naturally shows larger raw swings simply because it is larger.
In the third chart, we look at Absolute Percentage Error (APE). Here, the opposite problem appears: the low-volume series can suddenly look dramatically worse, because small raw misses create very large percentages when the denominator is small.
In the fourth chart, we look at square-root transformed error. Now the two series are much more comparable. The errors are not identical, but they are on a more balanced scale.
That makes it easier to compare forecast quality across skills without over-rewarding large skills or over-penalising small ones.
Summarising Performance: Transformed MAE
At the skill level, we usually care about more than one interval. What matters is the overall pattern over time. A simple and practical way to summarise this is to take the average of the transformed errors over all intervals. That gives:
This is simply the Mean Absolute Error (MAE), but calculated after applying the square-root transformation. This has a few practical advantages. Firstly, it is easy to explain and calculate. In addition, it gives one interpretable score per skill, making it more suitable for fair comparisons across skills than raw percentage-based metrics alone. A lower value means the forecasts are, on average, closer to actual arrivals on the transformed scale. A higher value means the forecasts are less accurate relative to the natural scale of that skill.
What This Means in Practice
This creates a useful two-level view of forecast performance. WAPE remains a strong headline metric for assessing overall forecast accuracy across the operation, and it is especially relevant for workforce planning because it preserves the operational importance of volume. At the same time, if the goal is to compare forecasting performance across individual skills, a transformed MAE can provide a fairer and more balanced basis for comparison by reducing the extent to which raw scale dominates the result. Used together, these measures allow you to retain a business-relevant view of total forecasting performance while also answering more diagnostic questions, such as which skills are genuinely harder to forecast, which are underperforming relative to their own scale, and which areas should be prioritised for forecast improvement, deeper analyst attention, model refinement, or operational review.
A Practical Use Case: Ranking Skills
One particularly useful application is ranking. If you rank skills by WAPE alone, low-volume skills may appear worse simply because percentages become unstable at small volumes. If you rank skills by raw absolute error alone, large-volume skills will almost always dominate. But if you rank skills using square-root transformed MAE, you get something more balanced: a list that is closer to âforecast difficultyâ or âforecast quality,â rather than just âforecast size.âÂ
A broader way to think about forecast accuracy
No single metric answers every forecasting question equally well. If the question is: âWas the center forecast good overall?â then WAPE may be exactly the right metric. If the question is: âWhich skills are least well forecasted in a fair, comparable way?â then a transformed error measure may be more informative. Those are different questions, and they may deserve different metrics.
As forecasting in contact centers becomes more advanced, the challenge is not only to ask whether the overall forecast was good, but also where it was less reliable, and why.
That is where broader forecasting thinking becomes valuable. Not because one new metric solves everything, but because better forecasting often starts with better questions.
At CCmath, we help contact centers explore those questions in practice. Through CCforecast, training, and consultancy, we support teams in building forecasting approaches that are not only accurate overall but also transparent, practical, and aligned with operational reality.
A practical next step is not necessarily to replace your current forecast KPI, but to add one extra comparison lens where it helps most: skill-level evaluation. In many operations, that alone can reveal useful improvement opportunities that remain hidden in aggregate reporting.
Next Steps
Weâd be happy to discuss how CCforecast, forecasting training, or tailored consultancy could support your planning process. We are also currently planning and working on implementing a square root transformation within the tool to help rank skills more effectively in terms of accuracy.