WFM guideForecasting

WFM forecasting methods

Q: What is the best forecasting method for contact centre WFM?

There is no single 'best' forecasting method for contact centre WFM; the right choice depends on your data volume, pattern characteristics, and operational context. As a practical guide: if you have fewer than 12 months of history and stable, seasonal volume patterns, weighted moving averages or Holt-Winters exponential smoothing are robust and require no statistical expertise. If you have 24+ months of data with complex seasonal patterns and identifiable trend, SARIMA models will typically outperform moving averages by 10–20% on WAPE. If you have 3+ years of data and known event drivers (marketing campaigns, product launches, billing cycles), regression-based models incorporating event variables are the most accurate for operations with strong event-driven volume. Machine learning approaches (gradient boosting, LSTM neural networks) offer marginal gains over SARIMA/regression for most contact centre applications unless you have very large datasets (millions of interval observations) and heterogeneous contact type mix. The most common mistake is choosing a complex model for insufficient data: a Holt-Winters model on 18 months of weekly data consistently outperforms an ARIMA model fitted to the same dataset, because ARIMA parameter estimation requires more data to be reliable.

Q: What is WAPE and how is it used in WFM forecasting?

WAPE (Weighted Absolute Percentage Error) is the primary accuracy metric for contact centre volume forecasting. It is calculated as: sum of |Actual − Forecast| ÷ sum of Actual, expressed as a percentage. Unlike MAPE (Mean Absolute Percentage Error), WAPE weights errors by actual volume, so a large error in a high-volume interval is weighted more heavily than the same absolute error in a low-volume interval. This is appropriate for contact centres because a 30% forecast error in a low-volume overnight interval has negligible operational impact, while a 15% error in a peak lunchtime interval drives significant SL impact. Turnella's forecast accuracy calculator computes WAPE from interval data. Industry benchmarks: WAPE below 10% at weekly level is considered good; WAPE below 5% at weekly level is excellent; WAPE below 15% at 30-minute interval level is considered acceptable; WAPE below 10% at interval level is excellent.

The algorithm is not the most important factor in forecast accuracy. Data quality and operational knowledge (knowing which events drive volume spikes, which channels substitute for each other, and which days are genuinely seasonal) account for more forecast error than the method. But the right method for the data pattern can reduce error by 10–30%.

Forecasting methods comparison

Method	How it works	Data requirement	Best for	Limitation
Simple moving average	Average of the last N periods. Equal weight to all periods in the window.	Minimum 4–8 periods. No pattern detection.	Stable volume with no trend or seasonality. Quick baseline for simple operations.	Lags behind trend changes. Does not capture seasonality.
Weighted moving average	Moving average where more recent periods receive higher weights. Weights are manually specified or decay by a set formula.	Minimum 4–8 periods.	Stable volume with slight trend where recent weeks are more representative than historical. Good when business is in a stable growth phase.	Weight selection is subjective. Does not handle multiple seasonal patterns.
Holt-Winters (triple exponential smoothing)	Separates the time series into level, trend, and seasonal components. Updates each component with separate smoothing parameters.	Minimum 2 full seasonal cycles (typically 2 years of weekly data). At least 52 weekly observations.	Most contact centres with clear weekly and annual seasonality. Standard method in enterprise WFM platforms. Good accuracy-to-complexity ratio.	Requires stable seasonality; poor for operations with changing seasonal patterns or large year-on-year structural changes.
ARIMA / SARIMA	Auto-Regressive Integrated Moving Average. Models time series as a function of its own lagged values, past errors, and seasonal components (SARIMA adds seasonal differencing).	Typically 100+ time periods for reliable parameter estimation. 2–3 years of interval data for interval-level SARIMA.	Operations with complex autocorrelation structures where volume in one interval depends on volume in previous intervals (e.g. contact types with repeat-contact patterns).	Requires statistical expertise for parameter selection. Sensitive to outliers. Does not naturally incorporate external variables (events, campaigns).
Regression (linear and event regression)	Models volume as a linear function of predictors: day of week, time of week, month, plus event dummy variables (bank holidays, marketing campaign live, billing cycle day).	Minimum 12 months for reliable coefficient estimation. Event calendar required.	Operations with strong event-driven volume (email bills, campaign responses, bank holiday patterns, billing cycles). Most practical for planners who can explain the drivers.	Linear relationship assumption may not hold. Requires maintained event calendar. Cannot capture complex non-linear patterns.
Machine learning (gradient boosting, LSTM)	Gradient boosting (XGBoost, LightGBM) treats forecasting as a supervised learning problem, learning non-linear relationships between features (calendar, lag features) and volume. LSTM (Long Short-Term Memory) is a neural network designed for sequential data.	Millions of interval observations for reliable training. Feature engineering expertise required.	Very large contact centres with heterogeneous contact type mix and complex seasonal interactions where linear models cannot capture the patterns.	High data and expertise requirement. 'Black box' models are difficult to explain to operational stakeholders. Marginal improvement over SARIMA/regression for typical contact centre datasets.

Method selection guide

Choose your forecasting method based on data volume and pattern complexity

If: Less than 12 months of history

Use: Weighted moving average or Holt-Winters with initial seasonal indices derived from business knowledge (e.g. 'December is our peak, estimate 1.4× average'). Complex models cannot be reliably fitted with insufficient data, so do not attempt ARIMA.

If: 12–24 months of weekly data, clear seasonal pattern, no major structural changes

Use: Holt-Winters (triple exponential smoothing). Available in Excel (FORECAST.ETS), most WFM platforms natively, and Python (statsmodels). Robust, explainable, and well-calibrated for most contact centre patterns.

If: 24+ months of interval data, stable operation, consistent contact mix

Use: SARIMA at interval level or Holt-Winters at daily level decomposed to interval using historical interval distribution. SARIMA offers marginal accuracy improvement over Holt-Winters for well-parameterised models.

If: Known event drivers (marketing campaigns, billing cycles, product launches, bank holidays)

Use: Event regression as primary method or as a correction layer on top of Holt-Winters. Add event dummy variables to a regression model. This is the most practically impactful enhancement for operations with event-driven volume: a marketing campaign that drives 20% more contacts is predictable with an event variable, invisible to a pure time-series model.

If: Large CC (500+ agents), 3+ years of data, complex multi-contact-type mix, dedicated analytics team

Use: Gradient boosting (XGBoost/LightGBM) with lag features, calendar features, and event variables. Expect 10–20% WAPE improvement over SARIMA at interval level for large, complex datasets. Do not attempt without statistical/ML expertise and significant feature engineering investment.

Accuracy benchmarks by horizon

Forecast horizon	Acceptable WAPE	Good WAPE	Excellent WAPE	Operational implication
Monthly (strategic resourcing)	<15%	<10%	<5%	Headcount sign-off and hiring decisions. A 15% error means hiring 15% too many or too few agents, a significant cost impact.
Weekly (scheduling)	<12%	<8%	<5%	Weekly schedule generation. 12% error at weekly level produces noticeable SL impact in peak intervals.
Daily (intraday planning)	<15%	<10%	<7%	Daily intraday allocation. Errors at daily level are partially absorbed by intraday management.
Interval (30-minute)	<20%	<15%	<10%	Real-time scheduling and break management. High interval-level variance is normal; the objective is to capture the pattern, not predict each interval perfectly.

Use the Turnella forecast accuracy calculator to compute WAPE and MAE from your interval data.

Forecasting method questions

What is the best forecasting method for contact centre WFM?

No single best method; the choice depends on data volume and pattern complexity. Under 12 months: weighted moving average. 12–24 months with seasonal pattern: Holt-Winters. 24+ months with complex patterns: SARIMA. Operations with event-driven volume (campaigns, billing cycles): event regression. Very large operations (500+ agents, 3+ years of data): gradient boosting with feature engineering. The most common mistake is using a complex model on insufficient data. Holt-Winters on 18 months outperforms ARIMA on the same dataset because ARIMA requires more data for reliable parameter estimation.

What is WAPE and how is it used in WFM forecasting?

Weighted Absolute Percentage Error: sum of |Actual − Forecast| ÷ sum of Actual. It weights errors by actual volume, so peak-interval errors matter more than overnight errors. Industry benchmarks: WAPE <10% at weekly level is good; <5% is excellent. WAPE <15% at 30-minute interval is acceptable; <10% is excellent. Use the Turnella forecast accuracy calculator to compute WAPE from your interval data.