Forecasting Methods
Forecasting Methods is the master reference for forecasting techniques used in workforce management. This page maps the method families, how they relate, and which to reach for when. Each family has its own dedicated page covering equations, parameter choices, common pitfalls, and WFM-specific application.
Forecasting in WFM is the demand side of the supply-and-demand operating model. Volume forecasts (calls, contacts, transactions), AHT forecasts, attrition forecasts, and capacity forecasts all share the same underlying methods — the difference is what's being predicted, not how.
When you need a forecast
The WFM team's forecasting workflow operates on multiple horizons:
- Long-range capacity planning (months to years) — annual budgets, hiring plans, site-level capacity
- Medium-range planning (weeks to months) — monthly forecasts, schedule generation, time-off allotments
- Short-range planning (days to weeks) — weekly volume forecasts, intraday distributions, surge plans
- Real-time (hours to minutes) — intraday updates, variance correction
Different horizons favor different methods. A daily call-volume forecast for next week and a quarterly attrition forecast use the same mathematical machinery but optimize for different things.
Method families at a glance
| Family | Best for | Complexity | WFM page |
|---|---|---|---|
| Naive / Seasonal Naive | Baselines, benchmarking | Trivial | Naive and Seasonal Naive Forecasting |
| Exponential Smoothing (ETS) | Trended/seasonal series with stable structure | Low–Medium | Exponential Smoothing |
| ARIMA | Series with autocorrelation requiring explicit modeling | Medium–High | ARIMA Models |
| Regression with explanatory variables | Forecasts driven by known business inputs (e.g., marketing spend) | Medium | Regression for Forecasting |
| Decomposition (STL, classical) | Diagnostic before model selection; interpretable trend/seasonal components | Low | Time Series Decomposition |
| Hierarchical reconciliation | Multi-skill / multi-site forecasts that must roll up consistently | High | Hierarchical Forecasting |
| Intermittent demand methods (Croston, SBA, TSB, ADIDA) | Series with frequent zero values (low-volume skills) | Medium | Intermittent Demand Forecasting |
| Forecast combination / ensembling | Improve accuracy across any methods you already run | Low | Forecast Combination |
| Probabilistic forecasting (PIs, quantile forecasts) | Risk-aware capacity planning; ranges instead of points | Medium | Probabilistic Forecasting |
| Judgmental forecasting (Delphi, scenario, adjustment) | Events not in historical data; SME knowledge layer | Variable | Judgmental Forecasting |
A decision tree for picking a method
- Always run a naive baseline first. If the fancier method cannot beat seasonal naive on the same data, the fancier method is broken. See Naive and Seasonal Naive Forecasting.
- Does the series have trend, seasonality, or both?
- Is there autocorrelation that ETS can't capture? Plot the residuals and the ACF. If significant autocorrelation remains after ETS, ARIMA is the next step.
- Are there known business drivers? (Marketing spend, product launches, seasonality from external events.) Regression with explanatory variables — combined with ARIMA errors as needed — captures driver-response patterns that pure univariate methods cannot.
- Multiple related series that must agree? (Skill-level forecasts that roll up to a site total.) Hierarchical forecast reconciliation.
The benchmark imperative
Across WFM contexts, the most common forecasting failure is not "we used the wrong method." It is "we used a sophisticated method without comparing it to the simple baseline." Hyndman's discipline is to compute the seasonal naive forecast on the same series and the same horizon and require any candidate method to beat it before adopting.
If your WFM software's forecast cannot beat seasonal naive on your historical data — and this happens — that is data telling you the method is misconfigured, the data is contaminated, or the method family is wrong for the series. The fix is diagnostic, not vendor escalation.
Forecast accuracy and evaluation
A method is only as good as its measured error. Common metrics:
- MAE — mean absolute error; in the units of the data
- RMSE — root mean squared error; penalizes larger errors more heavily
- MAPE — mean absolute percentage error; intuitive but unstable on near-zero values (common pitfall in skill-level WFM data with intermittent demand)
- MASE — mean absolute scaled error; the seasonal-naive-relative metric Hyndman recommends for cross-series comparison
See Forecast Accuracy Metrics for the full treatment. The headline rule: do not optimize MAPE on intermittent demand.
Time series concepts that show up in every method
- Trend — long-run direction of the series
- Seasonality — repeating pattern at a fixed period (daily, weekly, monthly, yearly)
- Cyclicality — non-fixed-period swings (business cycles)
- Noise — residual variation not captured by the structural components
- Stationarity — does the statistical structure of the series hold over time? Many methods (ARIMA in particular) require stationary input or differencing to achieve it.
- Decomposition — explicitly separating trend, seasonal, and residual components for diagnostic purposes
WFM-specific forecasting use cases
- Volume forecasting — calls, contacts, chats, emails by interval. The canonical WFM forecast.
- AHT forecasting — average handle time, often modeled separately from volume because drivers differ
- Attrition forecasting — separations by cohort or class. See the calculator pages Annual Attrition and Training Attrition for the ratio mechanics; method pages here handle the time-series methodology.
- Capacity forecasting — required FTE over horizon, derived from volume × AHT / occupancy with appropriate Erlang or simulation. See Demand calculation.
- Speed-to-proficiency curves — modeling how new hires ramp; relevant to Speed to proficiency curve.
Connecting forecasting to the WFM operating model
Forecasts are inputs to capacity planning, scheduling, and intraday operations. The Future WFM Operating Standard frames forecasting as a Level 2-and-above capability, with variance-aware operations requiring forecasts that express ranges rather than single numbers. The WFM Ecosystem Architecture places advanced capacity planning (Pillar 3) as the place where forecast distributions feed risk-aware decisions.
The methods on the linked pages are the math beneath those operating principles. Choosing the right method matters; using ranges rather than point estimates matters more.
References
- Hyndman, R. J., & Athanasopoulos, G. Forecasting: Principles and Practice (Python edition). otexts.com/fpppy. The canonical open-access reference for the methods on the linked pages.
See Also
- Naive and Seasonal Naive Forecasting — the baseline every method must beat
- Forecast Accuracy Metrics — MAE, RMSE, MAPE, MASE — which metric for which decision
- Regression for Forecasting — forecasts with known business-driver inputs
- Time Series Decomposition — STL and the diagnostic step before method selection
- Hierarchical Forecasting — multi-level forecast reconciliation (skill → site → enterprise)
- Intermittent Demand Forecasting — Croston, SBA, TSB, ADIDA for series with frequent zeros
- Forecast Combination — ensemble methods that beat any individual method
- Probabilistic Forecasting — prediction intervals and quantile forecasts for risk-aware planning
- Judgmental Forecasting — disciplined SME judgment layered on statistical methods
- Exponential Smoothing — the most-used family in WFM beyond Erlang
- ARIMA Models — when ETS isn't enough
- Demand calculation — calculator that consumes a forecast
- Power of One — interval-level sensitivity that depends on accurate forecasts
- Annual Attrition, Training Attrition — calculators that use forecasted attrition rates
- WFM Processes — where forecasting sits in the WFM operating model
- Future WFM Operating Standard — the strategic framework
- WFM Ecosystem Architecture — Pillar 3 (Advanced Capacity Planning)
