WFM Demand Forecasting Model

Interval-level intraday call volume forecasting, shrinkage modeling, Erlang C service-level estimation, and FTE gap analysis to support contact-center staffing optimization.

Total scheduled staffing landed within ~2% of daily requirement — yet the schedule left 53% of intervals understaffed and 44% overstaffed, holding delivered service level below the 80/20 target all day.

The headcount was right. The intraday distribution was wrong.

That gap — invisible to anyone reading only daily or weekly totals — is what interval-level WFM forecasting exists to catch. The aggregate looked fine on paper. The floor was understaffed more than half the day.

Method: profile-ratio interval forecast (21.2% WAPE on holdout) → Erlang C sizing to 80/20 → 34% shrinkage build-up → interval-by-interval FTE gap report. Real published arrival data (Technion SEE Lab, 87,600 rows); WFM overlay synthetic and labeled as such throughout.

This project demonstrates an end-to-end Workforce Management (WFM) forecasting and capacity-planning methodology in Python...

• Interval-level volume forecasting — a transparent profile-ratio (percent-of-day) forecast that separates the day-total expectation from the intraday distribution, validated on a holdout and reported with WAPE (the volume-weighted error metric WFM teams track), not just MAPE.
• Workload conversion — calls × Average Handle Time → offered load (Erlangs).
• Erlang C service-level modeling — sizing agents-on-phone to an 80/20 service-level target under a 90% occupancy ceiling, with the queueing assumptions and their real-world caveats stated plainly.
• Shrinkage modeling — a fully decomposed shrinkage build-up (breaks, training, coaching, PTO, unproductive time) uplifting the on-phone requirement to scheduled FTEs.
• Staffing gap analysis — required vs. a realistic lumpy shift schedule, interval by interval, surfacing over/under-staffed intervals and the delivered service level — demonstrating how total headcount can be right while the intraday distribution is wrong.

"Anonymous Bank" call-center dataset — Service Enterprise Engineering (SEE) Lab, Technion – Israel Institute of Technology (Prof. Avi Mandelbaum); documented by Guedj & Mandelbaum (2000) and widely used in the call-center / queueing literature (Brown et al., 2005).

Accessed via the Rfssa package data mirror, which publishes the 1999 arrivals pre-aggregated into 6-minute intervals (240 intervals/day × 365 days = 87,600 rows):

https://github.com/haghbinh/dataset/raw/main/Rfssa_dataset/Callcenter.rds

The notebook downloads this file automatically on first run.

My WFM experience is founder-era (a 350-agent call center, intraday management, service-level/abandonment SLAs, schedule adherence and shrinkage reporting) and predates the modern commercial suites (NICE, Verint, Calabrio). Those tools automate exactly the methodology shown here: interval forecasting, Erlang queueing, shrinkage uplift, and schedule-vs-requirement gap analysis. This project demonstrates that underlying methodology directly in Python — the reasoning that sits beneath any platform.

• The arrival data is real; the WFM overlay is synthetic and labeled as such throughout the notebook
• — Average Handle Time (240s), the 34% shrinkage build-up, the 80/20 service-level target, the 90% occupancy ceiling, and the shift schedule are explicit planning assumptions, not values derived from the data.
• Business logic, WFM domain framing, and operational assumptions are my own, drawn from contact-center workforce-management experience. AI tooling assisted with Python syntax and plotting code.

Python · pandas · NumPy · Matplotlib · Seaborn · pyreadr (to read the R .rds data) · Jupyter

pip install -r requirements.txt
jupyter notebook WFM_Demand_Forecasting.ipynb

Or open the notebook directly — every cell runs top-to-bottom and the dataset is fetched automatically.

Data source: SEE Lab, Technion (Prof. Avi Mandelbaum). The data is free for academic / non-commercial use; please acknowledge the source in any derived work.

Abandonment-aware queueing (Erlang A/X), intraday re-forecasting, multi-skill / blended workload, shift-bidding optimization, and a what-if sensitivity layer for AHT and shrinkage.