Call Center Outsourced evidence brief · Desk review · Published
What breaks when a forecast borrows another industry's seasonality curve?
A backtest plan for borrowed seasonality curves, including holiday calendars, contact mix, and interval-level error.

On this page
- Key stats
- Key takeaways
- Question and scope
- What a borrowed curve assumes
- Check the demand unit
- Compare holiday calendars
- Test the curve against local history
- Separate trend from season
- Model the contact mix
- Run a backtest
- Report error by interval
- Independent replay
- Limits and next step
- Decision the forecast supports
- Check model stability
- Involve the frontline
- Document the chosen model
- What would change the conclusion
- Review with staffing
- Publish the forecast note
- Put this into a support lane
- Related operating guides
- FAQs
- Sources
Key stats
- 2.67 million U.S. customer service representative jobs in 2025 (BLS)
- 5% projected decline in employment, 2025 to 2035 (BLS)
- 289,500 projected annual openings, mostly replacement (BLS)
Key takeaways
- Test a borrowed curve against a local baseline on a held-out period.
- Compare holiday calendars and contact units before fitting the curve.
- Report forecast error by interval and by day, not only as one annual average.
- Translate error into positions and wait time for the staffing decision.
Question and scope
This brief asks what happens when a call center forecast borrows a seasonality curve from another industry. The population is one queue's contact volume during a declared historical window. The unit is a forecast interval and its error. Freeze the interval, the horizon, and the error measure before testing the curve. Record the history available and the date of the last complete period. The test is whether the borrowed curve improves a staffing decision, not whether it looks plausible in a chart.
What a borrowed curve assumes
A borrowed curve assumes that the other industry's weekly shape, holiday effects, and seasonal peaks apply to this queue. That assumption may hold for a shared retail calendar and fail for a business with a different billing cycle. The curve is a hypothesis, not a neutral input. Write the assumptions down before fitting anything. A curve that cannot state its assumptions cannot be tested, because every miss becomes an exception after the fact.
Check the demand unit
Define whether the forecast is contacts, calls, tickets, or customer issues. A curve built from call volume may not fit a queue that receives most of its work through chat and email. Even within calls, a transferred contact and a new contact are different units. Use the same unit in the history and the forecast. A growth in transfers can look like demand growth when the customer issue count is flat. The unit should match the resource the forecast is meant to staff.
Compare holiday calendars
List the holidays that change the queue's volume and the days the business closes. Compare the borrowed curve's holiday assumptions with the local calendar. A curve that treats a local holiday as a normal weekday will overstaff the queue and misread the following day's recovery. Mark the holiday and the recovery day separately. The day after a closure often carries displaced demand, and a borrowed curve may have no equivalent event.
Test the curve against local history
Fit the borrowed curve to the local history and compare it with a simple local baseline. Use at least one full seasonal cycle if the data exists. If the borrowed curve does not beat the baseline on a held-out period, it is adding complexity without evidence. Keep the baseline simple enough to explain to an operations manager. A model that wins by a small margin but cannot be understood is a poor choice for a live staffing process.
Separate trend from season
A curve can fit the seasonal shape while missing a long-term change in volume. Detrend the history before comparing the seasonal pattern, then add the local trend back for the forecast. A queue that is growing or shrinking will otherwise attribute the trend to the season. Plot the trend and the seasonal index separately. The reviewer should see whether the borrowed shape survives after the trend is removed.
Model the contact mix
Forecast the total and the mix. A queue with a stable total but a changing mix can still miss its staffing plan, because different contact types have different handling times. Test whether the borrowed curve improves the total, the mix, or both. Build a separate error measure for each contact type. A curve that improves the total while making the complex queue worse is not an improvement for the customer who needs the complex queue.
Run a backtest
Hold out the most recent period and forecast it with the borrowed curve and with the local baseline. Compare mean absolute percentage error and the bias. Report the error by interval and by day of week, not only as one annual number. A curve that performs well on average can fail on the days that matter most. Include at least one holiday and one ordinary week in the test. The backtest should resemble the decisions the team will actually make.
Report error by interval
Show the forecast error for each interval and the staffing consequence. A ten percent error at a low-volume hour may be manageable, while the same error at peak may create a long wait. Translate the error into positions and wait time so the operations owner can judge the risk. Mark the intervals where the forecast changes the staffing plan. An error that does not change a decision is less urgent than one that does, even when the percentage looks similar.
Independent replay
Give a second analyst the same history, the same curve, and the same backtest rules. Ask for an independent result. Compare the forecast error and the treatment of holidays. Differences in the data cut-off or the holiday calendar should be resolved before the forecast is used for staffing. Record the resolution and the final parameters. The replay tests the forecast process, not the analyst's loyalty to the borrowed curve.
Limits and next step
The result depends on the history available, the stability of the queue, and the chosen error measure. It does not prove that the borrowed curve is always wrong. The next step is to keep the local baseline, document the borrowed assumptions, and revisit the choice after the next seasonal peak. Name the review date and the owner. A forecast choice that is never revisited becomes an unexplained tradition.
Decision the forecast supports
Use the backtest to decide whether to adopt the borrowed curve, blend it with local history, or keep the local model. Name the owner of the forecast and the date of the next review. A forecast is a decision aid, and its assumptions should be visible to the people who staff against it. Record the staffing change that the forecast is expected to produce. The test is complete when the team can explain why the chosen curve is better than the simpler alternative.
Check model stability
Re-estimate the model on a rolling window and watch whether the seasonal factors move. A stable curve is easier to defend than one that changes shape every month. If the factors move sharply, look for a change in the business before blaming the model. Record the estimation window and the last update. A forecast that is never refreshed will slowly become a description of an older queue, even when the code still runs without error.
Involve the frontline
Ask the supervisors and agents whether the forecast matches what they see. They know about local events, school calendars, and service outages that may not be in the data. Treat their observations as hypotheses to test, not as proof. A short weekly note can explain an unusual interval before it becomes a mystery in the monthly review. The forecast improves when the people who staff it can challenge its assumptions.
Document the chosen model
Write a short model note with the data window, the chosen curve, the holiday treatment, the error measure, and the next review date. Keep the note beside the forecast file so a new analyst can reproduce the result. Include the baseline and the reason it was rejected. A model that cannot be reconstructed is a personal artifact, not an operating tool. The note should be short enough to read before the staffing meeting.
What would change the conclusion
A new product launch, a change in the billing cycle, or a shift in channel mix can invalidate the borrowed curve. State those triggers and revisit the model when one occurs. Do not wait for the annual review if the business has already changed. A forecast is only as good as the stability of the process it describes, and the team should know when that stability has ended.
Review with staffing
Walk the forecast through the staffing plan before publishing it. Ask whether the interval-level error changes the number of positions or the start time of a shift. A forecast that is statistically better but operationally unusable is not an improvement. Record the staffing decision that depends on the forecast and the margin the manager uses. The review should include the supervisor who will have to explain the plan to the floor.
Publish the forecast note
Publish a short note with the chosen curve, the baseline, the error by interval, and the next review date. Keep it beside the forecast file, not in a separate folder that no one opens. The note should be understandable to a supervisor who does not build models. Include one sentence on what would cause the team to change the curve. A forecast that cannot be explained to the floor will be ignored when the queue gets busy.
Put this into a support lane
Bring one queue, its reporting extract, and the manager who owns the operating decision.
Review the queueRelated operating guides
FAQs
Does this study rank agents?
No. It describes a measurement design and the decisions the resulting evidence can support.
What counts as closure?
A reviewed result linked to the source, the definition, and the named owner who accepted it.