Objective
PMI Live Forecast Accuracy helps you review forecast values from selected snapshots alongside the comparison data shown in the report. It supports a practical question: how close was an earlier expectation to the outcome, and where should the forecasting process improve?
Use the view to investigate patterns across dates and departments. A single unusual day should prompt investigation rather than a conclusion about the reliability of every forecast.
Set the review context
Select the property or organizational unit, reporting period and snapshot date. The reporting period identifies the dates being reviewed; the snapshot identifies the forecast position used for the comparison.
Historical dates and future dates need different interpretation. An actual result is an outcome. A future Live Forecast remains an expectation. Check the column heading before describing a difference as forecast error.
Choose what to display
The legacy interface provides Revenue and Labour modes. Available measures include revenue, room nights or other applicable units, guest nights and ARR for revenue analysis, and hours and productivity for labour analysis.
Use View Options to choose data content and the measures required for the review. Options in the interface include Budget or Target, Forecast, Last Year, Live Forecast and Actual/Live Forecast, with availability depending on the view. Snapshot comparisons can use a Days before setting. Show dates adds date detail; MAD and MAPE can be displayed where applicable.
MAD expresses the size of deviations in the relevant measure’s units; MAPE expresses deviations as a percentage. Interpret them with the scope and underlying observations in mind. Low-volume or zero-actual periods need particular care.
Run a useful review
- Choose a completed period with sufficiently complete actual data.
- Select one department or a clearly defined group and one measure.
- Set the forecast snapshot or comparison horizon you intend to assess.
- Review the summary, then show dates to find where differences concentrate.
- Check import completeness and unusual operating events before attributing the difference to forecasting decisions.
- Agree a practical action, such as a regular review of event assumptions or earlier investigation of missing data.
Example
At Demo Hotel A, the team reviews room revenue for a completed week using an earlier forecast snapshot. Most days are close to the outcome, but one event date differs substantially. The team checks the booking changes and actual imports for that date before deciding how to improve the next review.
Common checks
If the result seems inconsistent, confirm the snapshot, period, department, measure and comparison content. Avoid comparing an earlier forecast for one report with today’s Live Forecast in another. When escalating a question, include these selections and the affected dates; share customer details only through the approved support channel.
Would the robot have been more accurate than my manual forecast?
To compare PMI Prediction (the robot) with a manual forecast fairly, use the values that were available at the same forecast horizon. Select the same property, dates, department and measure, and use sufficiently complete actual outcomes.
Obtain an archived automated prediction and manual forecast from matching lead times, such as both seven days before each stay date. A current or regenerated prediction may use information that was unavailable then and is not a valid substitute for that historical forecast.
The lead-time comparison matches each stay date with a saved forecast from the selected number of days before that date, then compares it with the actual outcome. Confirm which historical series the available report or export actually retains. A Live forecast snapshot does not automatically provide a separate archived robot prediction for every manually overridden day. If the required pair is unavailable, the comparison cannot be established from that report alone.
Compare error and bias on matching observations
Illustrative example using room nights, with both forecasts recorded at the same lead time:
| Stay date | Actual room nights | Archived robot forecast | Archived manual forecast |
|---|---|---|---|
| Day A | 100 | 90 | 95 |
| Day B | 80 | 90 | 100 |
The absolute errors are 10 and 10 for the robot, and 5 and 20 for the manual forecast. The mean absolute errors are therefore 10 and 12.5 room nights. On these two observations the robot has the smaller average absolute error. This small example does not establish that it is generally better.
Also inspect signed differences to see whether a forecast repeatedly overestimates or underestimates. Define the sign convention. If using a percentage error such as MAPE, state how zero or very small actuals are handled. Do not silently count a missing forecast as zero or compare different date coverage.
Use the finding to improve the review
Look for patterns in dates where manual adjustments helped or made the forecast less accurate, such as events, closures or unusual business. Check data quality and the evidence available at the time before judging the decision.
A lower error for one completed month is evidence about that period and horizon. It does not prove that every future date should be automated. Agree when to review or override the forecast and test the same comparison again over subsequent periods.