Forecasting · 10 min read

WFM Forecasting Models Explained

From basic averages to AI — the complete guide to choosing and evaluating forecasting methods for contact centres.

Why Forecasting Models Matter

A 5% forecast error on 1,000 calls per day means 50 calls go unplanned. At an average AHT of 5 minutes, that is over 4 hours of unscheduled workload — roughly 2 FTE for the entire day.

Forecasting accuracy is the single biggest driver of WFM efficiency. No amount of scheduling optimisation can compensate for a consistently inaccurate forecast.

Weighted Moving Average (WMA)

WMA is the most common baseline model used in contact centres. It averages historical data, with more recent periods given higher weight.

WMA = Σ(Weight × Historical Value) ÷ Σ(Weights) Example: 3-week WMA with weights 3, 2, 1: WMA = (3 × Week3 + 2 × Week2 + 1 × Week1) ÷ (3+2+1)

When to use WMA

Short-range forecasting (1–4 weeks ahead)
When recent data is more representative than older data
Simple, transparent, and easy to explain to stakeholders

Limitations

Does not model trend or seasonality — must be applied to deseasonalised data
Lags in detecting trend changes
Weight selection is manual and somewhat arbitrary

Simple Exponential Smoothing (SES) & Holt-Winters

Exponential smoothing assigns exponentially decreasing weights to older observations. Holt-Winters extends this to handle both trend and seasonality.

Holt-Winters (triple exponential smoothing): Level: Lₜ = α(Yₜ / Sₜ₋ₘ) + (1−α)(Lₜ₋₁ + Tₜ₋₁) Trend: Tₜ = β(Lₜ − Lₜ₋₁) + (1−β)Tₜ₋₁ Season: Sₜ = γ(Yₜ / Lₜ) + (1−γ)Sₜ₋ₘ Forecast: Ŷₜ₊ₕ = (Lₜ + h×Tₜ) × Sₜ₋ₘ₊ₕ α, β, γ = smoothing parameters (0 to 1)

When to use Holt-Winters

When data has a clear trend and seasonal pattern
Medium-range forecasting (1–6 months)
Built into most spreadsheet software (Excel, Google Sheets) and WFM tools

ARIMA

ARIMA (AutoRegressive Integrated Moving Average) is a more sophisticated statistical model that combines autoregression, differencing, and moving average components. SARIMA extends it to handle seasonality.

ARIMA(p, d, q) × SARIMA(P, D, Q)[m] p = autoregressive order d = integration/differencing order q = moving average order m = seasonal period (e.g. 7 for weekly, 30 for monthly)

When to use ARIMA

When you have sufficient historical data (>2 years recommended)
Complex seasonality patterns
Academic rigor required — ARIMA is statistically defensible
Requires statistical software (Python statsmodels, R)

Machine Learning Forecasting

ML models (XGBoost, Random Forest, LightGBM, LSTM neural networks) can capture non-linear patterns and incorporate many external variables. They tend to outperform statistical models on large, complex datasets.

Key advantages

Can incorporate external drivers: CRM data, marketing campaign flags, weather, product release dates
Non-linear pattern detection
Real-time reforecasting (update forecast as actuals come in during the day)

Key limitations

"Black box" — harder to explain to non-technical stakeholders
Requires large, clean datasets and data engineering skills
Risk of overfitting on historical anomalies
Higher operational complexity to maintain

Practitioner advice: Start with a well-calibrated Holt-Winters model and MAPE benchmarking. Only move to ML when you have identified specific patterns the statistical model consistently misses and have the data engineering capacity to support it.

Forecast Accuracy Metrics

MAPE — Mean Absolute Percentage Error

MAPE = (1/n) × Σ |Actual − Forecast| / Actual × 100 Example: Actual = 1,000, Forecast = 950 Error = |1,000 − 950| / 1,000 × 100 = 5%

Best for communicating accuracy as a percentage
Target: <5% weekly, <10% daily
Limitation: undefined when Actual = 0; inflated for near-zero actuals

MAE — Mean Absolute Error

MAE = (1/n) × Σ |Actual − Forecast| Example: If you average 50 calls off per interval, MAE = 50

Expressed in the same unit as the forecast (calls, emails, etc.)
More intuitive for operational stakeholders: "we are off by 50 calls on average"
Use when actual values can be zero or near-zero (MAPE breaks down)

Horizon	MAPE Target (Good)	MAPE (Acceptable)
Monthly	< 3%	< 7%
Weekly	< 5%	< 10%
Daily	< 8%	< 15%
Interval (30-min)	< 10%	< 20%

← Shrinkage Guide Schedule Efficiency →