Call Volume Forecasting Model

What it does

Projects historical contact volume into the future at day and interval (15/30 min) resolution and converts the forecast directly into capacity (FTE). The core backbone is Time-Series Decomposition: every volume series is split into level (base) + trend (directional growth/decline %) + seasonality (intraday profile, weekly profile, yearly season). The forecast is not a black-box single number but a transparent product: interval_forecast = base × trend × day_profile × week_coeff × season × special_day_multiplier.

When ≥2 full seasons of data exist the backbone becomes Holt-Winters Triple Exponential Smoothing (multiplicative) — updating level/trend/season via α, β, γ; the multiplicative form fits call volume because seasonal effect scales with level. For short/sparse series it falls back to a seasonal-naïve baseline and flags this explicitly. The forecast is delivered as a confidence band (low/high scenario), and the WFM plan is built off the upper edge.

The critical step is the Erlang-C Capacity Bridge: traffic intensity A = (λ × AHT)/3600 Erlangs is computed, the smallest agent count N meeting the SL target (e.g. 80/20) is selected, then shrinkage inflates it to rostered_FTE = N / (1 − shrinkage). Because Erlang-C grows over-optimistic above ~90% occupancy, the model raises an abandon/burnout risk warning.

When to use it

When you hold at least a few weeks (ideally 13+ months) of historical volume and need an interval-based call forecast plus the staffing (FTE) it requires for the coming day/week/month. It goes beyond "how many calls" to "how many agents per interval, will I hit SL, how do I inflate the roster for a campaign day." Built for WFM/Operations teams, schedulers, and capacity decision-makers.

Method / frameworks

• Holt-Winters (Triple Exponential Smoothing) — industry-standard seasonal forecasting (level/trend/season).
• Erlang-C / Erlang-A traffic theory — mathematical basis of capacity planning; converts forecast to FTE.
• WFM discipline chain (Forecast → Schedule → Adherence → Shrinkage) and the COPC CX Standard operational frame.
• MAPE / WAPE accuracy metrics — measured via walk-forward backtest; daily benchmark of ~±5% is good, interval level far looser.
• Croston / intermittent demand split — low-volume channels are flagged and aggregated daily/weekly instead of by interval.
• Data-sufficiency gate — method and confidence auto-selected by history length; no fake confidence.