WFM guideForecasting

Contact centre forecast accuracy benchmarks

Q: What is a good forecast accuracy (WAPE) for a contact centre?

It depends entirely on the level of aggregation, and any benchmark that ignores that is misleading. At daily volume level, a WAPE (Weighted Absolute Percentage Error) of under 5% is strong, 5-10% is good and typical for a stable operation, and above 15% suggests a forecasting problem worth investigating. At weekly level, accuracy is usually a little better again (more aggregation). At interval level (e.g. 30-minute buckets), the same well-run operation might show 15-25% WAPE — and that is normal, not a failure, because each interval carries far less volume and random arrival variation is a much larger fraction of the total. The key principle: accuracy degrades as granularity increases, because aggregation cancels out random noise. A forecast that is 6% WAPE daily but 22% WAPE at interval level is performing normally. Channel and volume also matter: high-volume voice queues forecast more accurately than low-volume specialist queues; stable demand forecasts better than event-driven or campaign-driven demand. Rather than chasing a single number, judge accuracy against the level you measure at, track the trend over time (is it improving?), and watch for systematic bias (a consistent direction of error) separately from the size of the error.

Everyone wants to know if their forecast is "good". The honest answer depends on the level you measure at: daily accuracy is always better than interval accuracy. A daily WAPE of 5% is strong; the same operation may run 15–25% at interval level — and that is normal, not a failure.

Why accuracy depends on the level you measure

Forecast accuracy degrades predictably as the measurement gets more granular, because aggregation cancels random noise. A daily total pools every interval's random arrival variation, so the highs and lows offset and the total is stable and forecastable. A single 30-minute interval has none of that cancellation — the random arrival variation is a large fraction of the small volume, so the percentage error is inherently higher. This is not a forecasting weakness; it is statistics. The corollary: never quote a forecast accuracy number without stating the level it was measured at. "Our WAPE is 7%" is meaningless without "…at daily level" or "…at interval level" — the same forecast can be both.

Realistic WAPE benchmarks by level

Level	Strong	Typical	Investigate
Daily volume	< 5% WAPE	5–10% WAPE	> 15% WAPE
Weekly volume	< 4% WAPE	4–8% WAPE	> 12% WAPE
Interval (30-min)	< 12% WAPE	15–25% WAPE	> 35% WAPE

Indicative ranges for a stable, medium-to-high-volume operation. Treat them as orientation, not a contractual target — your achievable accuracy depends on your channel, volume, and how event-driven your demand is.

Daily volumeA full day aggregates all intervals, cancelling most random arrival noise. The daily total is driven by predictable factors (day of week, trend, events), so it forecasts well.

Weekly volumeEven more aggregation than daily. Weekly totals are the most stable and forecastable level — useful for capacity planning, less useful for scheduling.

Interval (30-min)Each interval carries little volume, so random arrival variation is a large fraction of the total. High interval WAPE is normal — it does not mean the forecast is bad. This is the level that actually drives staffing, so it matters most despite being least accurate.

What else affects achievable accuracy

Volume

Higher-volume queues forecast more accurately at every level — more volume means random variation is a smaller fraction of the total. A low-volume specialist queue will never match a high-volume queue's WAPE, and shouldn't be held to the same number.

Channel

Stable, habitual channels (routine voice/service) forecast better than reactive ones. Outbound campaigns, event-driven spikes, and new digital channels are inherently harder to forecast — expect higher WAPE and don't treat it as failure.

Demand stability

Operations with steady, seasonal, habitual demand forecast well. Operations dominated by one-off events, marketing campaigns, incidents, or external shocks have higher irreducible error — the demand itself is less predictable.

Bias is separate from accuracy

A forecast can be accurate (low WAPE) but biased (systematically over or under), or unbiased but inaccurate (large random error in both directions). Track bias separately — a small but systematic bias is more fixable, and often more damaging to staffing, than a larger random error.

Why chasing a single accuracy number is the wrong goal

A fixation on one headline accuracy figure leads to the wrong behaviours. Better questions than "what is our WAPE?":

→Is accuracy improving over time? The trend matters more than any single month's figure.
→Is there systematic bias? A consistent over- or under-forecast is more actionable (and often more harmful) than random error of the same size.
→Where is the error concentrated? A good daily WAPE hiding a terrible specific day-of-week or interval is the real problem.
→Is the accuracy good enough for the decision it feeds? Capacity planning needs daily/weekly accuracy; scheduling needs interval accuracy. Judge each against its purpose.

Forecast accuracy benchmark questions

What is a good forecast accuracy (WAPE) for a contact centre?

It depends on the level of aggregation. At daily volume level, WAPE under 5% is strong, 5–10% is good and typical, and above 15% warrants investigation. At weekly level, accuracy is usually a little better again. At interval level (30-min), the same well-run operation might show 15–25% WAPE — and that is normal, because each interval carries little volume so random arrival variation dominates. Accuracy degrades as granularity increases because aggregation cancels noise. A forecast that is 6% WAPE daily but 22% at interval level is performing normally. Channel and volume matter too: high-volume stable queues forecast better than low-volume or event-driven ones. Rather than chasing one number, judge accuracy against the measurement level, track the trend, and watch systematic bias separately from error size.