Skip to main content
TurnellaBeta
WFM guideForecasting

Contact centre forecasting review

A forecasting review is not a post-mortem on a bad month — it is a structured process for improving the model. The goal is not to explain away the errors but to understand them well enough to reduce them systematically. Most forecast errors have identifiable, correctable causes.

WAPE benchmarks by time granularity

WAPE (Weighted Absolute Percentage Error) is the primary accuracy metric for contact centre volume forecasting. WAPE naturally increases as time granularity increases — interval-level forecasting is inherently less accurate than monthly forecasting because individual interval volume is more variable.

GranularityExcellentAcceptableIndicates model problems
Monthly<5%5–8%>10%
Weekly<7%7–12%>15%
Daily<10%10–15%>20%
30-minute interval<15%15–25%>30%
Volume accuracy is not the same as staffing accuracy: A centre with 5% monthly WAPE and 35% interval WAPE has a distribution model problem. The total volume forecast may be accurate, but if volume is distributed incorrectly across the day (e.g. all volume assigned to the morning when afternoon is actually larger), the schedule will be wrong even though the monthly total is right. Track WAPE at every level separately.

Classifying forecast errors: systematic vs. random

The first task of a forecasting review is to classify each error as systematic (a correctable model or process failure) or random (inherent variance that cannot be eliminated through model improvement). Only systematic errors warrant model changes. Attempting to eliminate random errors by over-fitting the model to historical anomalies will reduce future accuracy, not improve it.

Systematic errors — correctable

Seasonal pattern error

Example

Forecast assumed December volume would be 30% below November based on prior years; actual December was only 15% below due to product launch

Model change / fix

Update the seasonal index for December. If a product launch is planned, this should be treated as a forecast event overlay — additive to the seasonal baseline, not replacing it.

Trend underestimation

Example

Forecast assumed flat volume trend; actual volume grew 8% month-on-month for three consecutive months due to customer base growth

Model change / fix

Incorporate a growth trend component into the forecasting model. Review the trend assumption quarterly and update when the 3-month rolling average shows sustained divergence from the prior year trend assumption.

Day-of-week pattern drift

Example

Historical pattern showed Monday volume 140% of weekly average; actual Monday volume has shifted to 120% over the past 6 months as customer behaviour has changed

Model change / fix

Recalculate day-of-week indices using only the most recent 12 weeks of data. Day-of-week patterns can drift significantly — using 12 months of historical data may be masking a recent structural shift.

Unforecast campaign or event

Example

Marketing sent an email campaign on a Thursday that was not communicated to the WFM team — actual volume was 40% above forecast for 3 hours

Model change / fix

This is a process failure, not a model failure. Implement a notification SLA: marketing must notify WFM at least 5 working days before any customer communication expected to generate contact.

AHT shift

Example

Volume forecast was accurate; SL was worse than expected because AHT increased from 360 to 430 seconds due to a new product query type

Model change / fix

AHT changes are a separate forecasting input from volume changes. Update the AHT assumption in the staffing model. Investigate whether the new product query type requires a knowledge base update to reduce AHT going forward.

Random errors — inherent

Random volume variance

Example

Forecast for a midweek day was accurate to 3%; actual volume on that day was 18% above forecast with no identifiable external cause

Response

No model change required. Inherent random variance — the irreducible uncertainty in customer behaviour. Managed through capacity buffer and intraday management, not through model improvement.

The monthly forecasting review: structure

1.

Calculate WAPE at monthly, weekly, daily, and interval level

Do not rely on a single WAPE figure. A good monthly WAPE can mask poor daily or interval accuracy. Calculate WAPE separately for each channel (voice, email, chat) and each time granularity. Flag any level where WAPE exceeds the acceptable threshold.

2.

Identify the largest absolute errors and classify each as systematic or random

Sort intervals, days, or weeks by absolute volume error (actual minus forecast). Focus review time on the largest errors — the 20% of intervals that account for 80% of the total error. For each large error, determine whether there is an identifiable cause (systematic) or no identifiable cause (random).

3.

For each systematic error, identify the root cause and the model change required

Common root causes: seasonal index outdated, trend assumption wrong, campaign not communicated, new contact type not in the model, AHT assumption shifted. For each, document the specific model change to apply in the next forecast cycle.

4.

Update the assumptions log

The assumptions log is the written record of every assumption embedded in the forecasting model — trend rate, seasonal indices, day-of-week distribution, AHT by contact type. Update any assumption that the review has shown to be wrong. Include the date of update and the evidence that drove the change.

5.

Set a WAPE improvement target for the next review

If current monthly WAPE is 12% and the acceptable target is 8%, the next review should target 10% (a realistic improvement step, not a target that requires perfect forecasting). If current WAPE is already within the acceptable range, the target is maintenance — do not introduce model complexity to try to achieve 'excellent' when 'acceptable' is already delivering operationally.

Forecasting review questions

What is a good WAPE for a contact centre volume forecast?

WAPE benchmarks vary by granularity. At monthly level: below 5% is excellent, 5–8% acceptable, above 10% indicates model problems. At weekly level: below 7% excellent, 7–12% acceptable, above 15% indicates problems. At daily level: below 10% excellent, 10–15% acceptable, above 20% indicates problems. At 30-minute interval level: below 15% excellent, 15–25% acceptable, above 30% indicates problems. A centre with good monthly WAPE but poor interval WAPE has a volume distribution model problem — the total is correct but the intraday shape is wrong.

Related guides