Why contact centre forecasting matters more than ever
Leaders need reliable service levels without waste. Planners need staffing numbers they can defend. Customers want fast, fair access. Forecasting converts volatile arrivals into the staffing plan that keeps wait times predictable at sustainable cost. Queueing science shows wait time rises sharply as utilisation nears full capacity and as variability increases, so even small forecast errors can create large swings in waits and abandons.¹ High-performing centres forecast at two horizons: a long-horizon plan for hiring and shrinkage, and intraday reforecasts that steer breaks, callbacks, and cross-queue moves. When forecasts are accurate and refreshed, service stabilises, cost per contact falls, and agent wellbeing improves.²
What are we forecasting and how do we define the workload?
Teams forecast arrivals by interval and convert arrivals into workload by multiplying by average handle time, then adjust for concurrency and after call work. Planners separate the statistical forecast from the policy translation that adds shrinkage and occupancy targets.³ Centres should treat each intent and channel as its own series when volumes allow, then reconcile to the total so operations numbers add up. Hierarchical reconciliation improves total accuracy without hiding local patterns.⁴ Good definitions keep the math clean and the plan auditable.
What data and features move accuracy the most?
Analysts start with clean interval data and clear calendars. They encode day-of-week, week-of-year, month, holidays, billing cycles, school terms, marketing events, outages, and product releases. Call centre research shows strong intraday seasonality and calendar effects with overdispersion around Poisson baselines, so feature-rich models beat naive repeats.⁵ Teams add digital leakage signals such as portal outages and login error rates because these drive sudden voice spikes. They track policy changes and pricing events as regressors. Data hygiene matters more than exotic algorithms. When event logs are reliable and features are honest, even simple models win.⁶
Which forecasting methods consistently work in contact centres?
Analysts keep a small, proven toolbox and use it well.
Exponential smoothing (ETS) handles level, trend, and seasonality with robustness and speed for hundreds of time series.⁷
ARIMA/SARIMA models capture autocorrelation and seasonal effects where residuals still show structure after ETS.⁶
Regression with ARIMA errors adds external drivers like promos, outages, or bill runs while controlling for serial correlation.⁶
Additive models such as Prophet provide fast, explainable fits for business calendars with holidays and changepoints.⁸
Gradient boosting or random forests add nonlinear interactions when you have rich exogenous features, though they need careful cross-validation to avoid overfitting.⁶
Ensembles that average top performers often beat any single model. The M4 competition confirms that simple, diverse combinations are hard to beat at scale.⁹
Teams favour methods that are fast to retrain hourly and transparent enough for operations to trust.
How do you forecast intraday patterns reliably?
Operations need interval shape, not just daily totals. Analysts model two layers. First, forecast the daily total with ETS or ARIMA plus regressors. Second, allocate the total into intraday shares using historical shape profiles or a second-stage model that predicts each 15- or 30-minute proportion by calendar features and recency.⁵ This two-stage approach is robust to special events and gives intraday managers a lever to rescale shape when real-time arrivals drift.
What accuracy measures and testing regimes keep you honest?
Planners use holdout periods and rolling-origin cross-validation rather than a single split. They report WAPE or sMAPE for scale-free accuracy, and MASE to compare against a naive seasonal benchmark.⁶ They add prediction intervals at 80 and 95 percent so intraday can plan buffers. M4 results caution against chasing a single perfect metric. Use at least two complementary measures and a naive baseline.⁹ Stability and error distribution matter as much as headline error.
How do you blend top-down and bottom-up without double-counting?
Hierarchical and grouped reconciliation lets you forecast at multiple granularities and then reconcile to a coherent whole. You can model intent, channel, and site separately, then reconcile with MinT or similar methods so child series sum to parents while preserving as much forecast signal as possible.⁴ This produces operational detail without losing the total staffing truth your CFO needs.
How should you handle special events, outages, and shocks?
Modelers build an event library with start time, end time, and expected uplift type. They create binary or ramp features for events and decay functions for after-effects. After unusual spikes, they downweight those intervals during model training so the future is not haunted by a one-off. For outages and incidents, teams layer a nowcast that blends the statistical forecast with live arrival rates to steer immediate actions like callbacks and break moves.⁵ Event hygiene is the difference between a forecast you can explain and a number you cannot defend.
What is the intraday playbook once the day starts?
Intraday teams run a closed loop of observe, compare, act. They compare live arrivals and handle time to forecast, compute variance by interval, and choose the lightest correction first: move breaks, switch on virtual-hold callbacks, shift concurrency, and reassign skills. When variance persists, they reforecast the remaining day using the same model stack and the observed morning shape. Research on callback policies confirms that offering callbacks above defined queue thresholds reduces abandonment and perceived wait while smoothing peaks.¹⁰ Forecast plus playbook is what turns numbers into service.
How do you translate forecast to staffing without surprises?
Planners convert workload to required staffing with Erlang C or simulation, then add shrinkage and set occupancy targets by interval.³ They publish assumptions about leave, training, coaching, meetings, and unplanned absence, then reconcile actuals monthly. They model uncertainty by staffing to the forecast plus a percentile of the error distribution for critical intervals. Because wait time explodes near saturation, buffers are cheaper than chronic overtime and burnout.¹
What are the pitfalls that quietly degrade forecast performance?
Five errors recur. Teams forecast totals and ignore intraday shape. Teams skip exogenous drivers and then blame “unpredictability.” Teams measure accuracy at the day level and miss interval pain. Teams chase a single algorithm and neglect ensembles that generalise better. Teams do not version their features and cannot reproduce results. Good forecasting practice fixes these with shape models, regressors, robust metrics, humble ensembles, and simple MLOps that version data, code, and parameters.⁶ ⁹
How to start or reboot your WFM forecasting in 30 days
Week 1. Clean 18 months of interval data. Encode calendar and event features. Establish naive and ETS baselines.⁶ ⁷
Week 2. Add ARIMA with regressors and a daily-total plus intraday-shape model. Produce rolling cross-validation and WAPE, sMAPE, and MASE.⁶⁵
Week 3. Create a small ensemble and generate prediction intervals. Reconcile across intent, channel, and site.⁴ ⁹
Week 4. Operationalise. Publish a forecast pack with intervals, intervals’ error bands, and a playbook of intraday actions at defined variance thresholds. Switch on hourly reforecasting and callbacks at queue thresholds.¹⁰
What impact should executives expect from better forecasting
Expect fewer peak-hour service misses, lower abandonment, and tighter adherence to planned staffing. Expect fewer emergency overtime requests and more predictable occupancy. Expect improved agent wellbeing because peaks hurt less and coaching windows survive. Forecasting that you can explain and refresh becomes a strategic asset because it lets you decide, not react.
FAQ
Which single method should we start with if we are new to forecasting?
Start with exponential smoothing for each series and add calendar features. It is fast, accurate, and easy to maintain. Layer ARIMA with regressors when you need external drivers.⁶ ⁷
How often should we reforecast intraday?
Reforecast at least hourly, or at each significant variance threshold. Blend statistical forecasts with live arrival rates to steer callbacks and break moves.⁵ ¹⁰
Do machine learning models beat classical methods in contact centres?
Sometimes, but not reliably without strong features and cross-validation. The M4 competition showed simple or hybrid ensembles often outperform complex single models.⁹
How do we forecast new queues with little history?
Borrow shape from similar queues, use short-term features, and apply hierarchical reconciliation so parent accuracy props up the child. Update quickly as data accrues.⁴ ⁶
Which accuracy metric should we report to the business?
Report WAPE or sMAPE for scale-free error and include prediction intervals. Add interval-level accuracy so intraday pain is visible. Use MASE to compare against a naive seasonal baseline.⁶
How do we connect forecasts to staffing decisions transparently?
Publish the chain: arrivals to workload, workload to Erlang or simulation staffing, staffing to shrinkage and occupancy, plus buffers based on forecast error percentiles.³ ¹
Sources
Queueing theory and service variability (Kingman’s formula overview) — Wikipedia contributors, 2025, Wikipedia. https://en.wikipedia.org/wiki/Kingman%27s_formula
Workforce Management Best Practices — NICE, 2024. https://www.nice.com/resources/workforce-management-best-practices
Call Center Staffing and Scheduling — Call Centre Helper Guide, 2023. https://www.callcentrehelper.com/call-centre-staffing-132.htm
Forecasting: Principles and Practice (chapters on hierarchical and grouped forecasting) — Rob J. Hyndman, George Athanasopoulos, 3rd ed., OTexts, 2021. https://otexts.com/fpp3/
Forecasting Call Center Arrivals: Methods and Models — Lawrence V. Brown, Noah Gans, Avishai Mandelbaum, Anat Sakov, Haipeng Shen, Sergey Zeltyn, Linda Zhao, 2005, Manufacturing & Service Operations Management. https://pubsonline.informs.org/doi/10.1287/msom.1050.0073
Forecasting: Principles and Practice (ETS, ARIMA, regression, evaluation) — Rob J. Hyndman, George Athanasopoulos, 3rd ed., OTexts, 2021. https://otexts.com/fpp3/
Exponential Smoothing: The state of the art — Everette S. Gardner Jr., 1985, Journal of Forecasting. https://onlinelibrary.wiley.com/doi/abs/10.1002/for.3980040103
Forecasting at scale — Sean J. Taylor, Benjamin Letham, 2018, The American Statistician. https://www.tandfonline.com/doi/full/10.1080/00031305.2017.1380080
The M4 Competition: Results, findings, conclusion and way forward — Spyros Makridakis, Evangelos Spiliotis, Vassilios Assimakopoulos, 2018, International Journal of Forecasting. https://www.sciencedirect.com/science/article/pii/S0169207018300787
Optimal scheduling in call centers with a callback option — Benoît Legros, 2016, European Journal of Operational Research. https://www.sciencedirect.com/science/article/abs/pii/S0166531615000930