Definition

A forecasting model uses historical demand to estimate future volume for each queue, channel, and time interval. ZenOps WFM can compare several models against actual historical volume so planners can see which predictions fit the data best.

Context

Open a forecast’s Statistics view to compare model results for a queue and channel. The Current badge marks the model being used for that queue. Select Set current on another model to use it instead, or select Show all models to expand the comparison.

Each fold trains on earlier demand and predicts a held-out period from known history. The dashed blue prediction is compared with the dark actual line; repeating this across folds helps evaluate a model before forecasting future demand.

Models are evaluated on historical periods held out from training. Each cross-validation fold is a separate period; the fold count and range of results help show whether a model performs consistently. These results describe past periods and do not guarantee the same accuracy in the future.

The Forecasts and Actuals chart overlays model predictions and observed volume for an evaluation period. Compare the lines across peaks, dips, and quieter intervals rather than focusing on a single day. A model’s fit can differ across queues, so review the result for the queue you are planning before setting its current model.

Models Used

The model comparison includes:

  • N-Week Average: Averages the same time interval across the selected number of previous weeks. It can work well when demand follows a steady weekly pattern.
  • Theta: Combines a trend estimate with a smoothed view of demand and its seasonal pattern.
  • Seasonal Naive: Uses the corresponding interval from the previous weekly cycle. It is the locked baseline for comparing other models.
  • Holt-Winters: Models the level, trend, and seasonal pattern in the historical demand.
  • Ensemble: Combines the Seasonal Naive, N-Week Average, and Holt-Winters forecasts into a weighted blend. The weights reflect how those component models performed on historical validation data.
  • TimesFM 2.5: Uses Google’s time-series foundation model to generate a forecast; its results are evaluated with the same accuracy measures as the other candidates.

Accuracy Metrics

Use these measures to compare model error and understand its direction and practical effect:

  • Accuracy: A simplified 0–100 score calculated as max(0, 100 − WAPE). It is not the percentage of intervals predicted exactly.
  • WAPE: Total absolute forecast error as a share of actual volume across the evaluation periods. Lower WAPE indicates less overall error and is a clear starting point for comparing models.
  • MAPE: Average percentage error for intervals with non-zero actual volume. Quiet intervals can have a large effect on this measure, so it may tell a different story from WAPE.
  • Over: Average contacts and percentage error in intervals where the model predicted too much volume.
  • Under: Average contacts and percentage error in intervals where the model predicted too little volume.
  • Bias: The average direction of the error. Positive bias indicates over-forecasting; negative bias indicates under-forecasting.
  • Direction: How often the model correctly predicts whether volume will rise or fall in the next interval. Long stretches of zero volume can make this measure look stronger than the forecast is for a sparse queue.

To investigate a forecast that does not match actual demand, see Investigate Inaccurate Forecasts. To review demand and coverage together, see How To Review Forecast Outcomes.

Example: Choosing a model for a Level 1 baseline
Example results from a Level 1 baseline forecast comparison. Three of the six models are visible.

Reading the result. Among the visible candidates, TimesFM 2.5 is current and has the highest Accuracy (76.4%) and lowest WAPE (23.6%). It is the strongest visible choice when overall volume error is the priority. The screenshot shows three of six models, so compare all candidates before changing the current model.

Metric trade-offs. Theta has lower MAPE (34.2% versus 37.6%) and higher Direction (74.3% versus 66.5%). If percentage error per non-zero interval or predicting whether demand rises or falls matters more for your queue, inspect those results too. The right choice depends on the planning need.

Forecasting. With TimesFM 2.5 selected, future intervals use its volume estimates for that queue and channel. Its −1.4 bias here indicates a modest under-forecasting tendency across the evaluation data. Treat that as a signal to review staffing exposure around peaks, not as a guaranteed shortfall in every interval.

Reforecasting. After rerunning with updated history or assumptions, compare the new forecast with the baseline for the same queue, channel, and horizon. Once actuals are available, review errors on completed intervals. Rankings can change as new data arrives, so review coverage updates before relying on them.

Scheduling. Forecast volume can inform required-coverage calculations and schedule generation. A generated schedule starts as a draft: check coverage, constraints, and known events, then save and publish when it is ready. Choosing a model does not publish a schedule.