# Investigate Inaccurate Forecasts

## Issue(s)

Forecasted inbound volume differs from actual demand, the selected model repeatedly over- or under-forecasts, or model accuracy changes substantially across evaluation periods.

## Environment

- You have a completed forecast with actual historical demand available for comparison.
- You can open the forecast’s **Statistics** view for the affected queue and channel.

## Cause(s)

- The forecast and actuals use different date ranges, timezones, queues, channels, or intervals.
- Historical demand is incomplete, incorrectly mapped, too short, or no longer representative of normal weekly patterns and seasonality.
- An exceptional event is included in the history, a recurring peak was excluded, or a known future demand change is not represented.
- The current model does not fit the queue’s demand pattern, or its results vary across the historical evaluation periods.
- A metric is being read without its context: MAPE can be sensitive to low-volume intervals, and Direction can look strong when many intervals have zero volume.

## Diagnostic

1. In **Statistics**, confirm the queue, channel, date range, interval, and timezone. Compare the forecast with actuals from the same scope and period.
2. Check whether historical data is complete and mapped to the intended queues and channels. Investigate missing intervals or gaps with [Find Missing Forecast Data](/help/articles/find-missing-forecast-data/).
3. Review the historical period and interval in **Parameters**. Confirm that the history reflects normal weekly patterns and expected seasonality, and that enough evaluation folds are available for a useful comparison. See [How To Change Forecast Parameters](/help/articles/change-forecast-parameters/).
4. Inspect unusual spikes in the history. Exclude only exceptional periods that should not inform future demand; keep recurring peaks. See [How To Configure Outliers](/help/articles/configure-outliers/). For a known future campaign or event, check whether a demand adjustment represents it; see [How To Configure Adjustments](/help/articles/configure-adjustments/).
5. Compare candidate models over the same evaluation periods. Use WAPE to compare overall error and the fold range to assess consistency. Interpret MAPE carefully for low-volume queues, and review **Over**, **Under**, and **Bias** to identify the direction and operational effect of errors. Use **Direction** as additional context, especially when the queue has many zero-volume intervals. See [Understanding Forecast Models And Accuracy](/help/articles/understanding-forecast-models-and-accuracy/) for metric definitions.
6. In **Forecasts and Actuals**, compare each model’s line with actual demand across peaks, dips, and quieter periods. Check whether the model marked **Current** is the one whose pattern best fits the queue’s planning needs; a single lowest error score may not account for the operational cost of repeated under- or over-forecasting.

After identifying the cause, correct the data or settings, select an appropriate model if needed, and rerun the forecast to review the updated comparison.

> Canonical page: https://zenops.cx/help/articles/improve-forecast-accuracy/
> Audience: For planners and managers
> Summary: Diagnose forecast mismatches by checking comparison scope, historical data, parameters, and model results.

## Notes on this article

### Note: Volume and coverage have separate causes

In diagnostic. A better volume forecast does not automatically correct staffing assumptions. If inbound volume looks right but required coverage does not, review the queue’s forecasting profile.