THE RESULT
What the data shows
The model underpredicts in 63.8% of observed hours despite an 11.5% overall MAE improvement. Forty of 2,928 evaluation-calendar hours have no observed target.
On working-day evening commute hours, model MAE is 123.63 versus 121.31 for the baseline: 1.9% worse despite the overall improvement. This exploratory slice shows why the aggregate score is insufficient for planning.
| Month | observed hours | rentals | model mae | baseline mae | bias | wape | underprediction share | underprediction amount | mae improvement | expected hours | missing hours | coverage |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2012-09 | 720 | 218,573 | 68.57 | 61.75 | -43.19 | 0.2259 | 0.6778 | 40,234.56 | -0.1104 | 720 | 0 | 1 |
| 2012-10 | 708 | 198,841 | 76.01 | 67.09 | -55.17 | 0.2707 | 0.7669 | 46,437.78 | -0.1329 | 744 | 36 | 0.9516 |
| 2012-11 | 718 | 152,664 | 51.35 | 78.31 | -9.352 | 0.2415 | 0.546 | 21,792.86 | 0.3442 | 720 | 2 | 0.9972 |
| 2012-12 | 742 | 123,713 | 45.79 | 65.3 | -6.09 | 0.2746 | 0.5647 | 19,247.14 | 0.2988 | 744 | 2 | 0.9973 |
Inside the Power BI project
Three editable report pages, 17 DAX measures, and a documented model with active single-direction relationships.
- Forecast reliability
- Underprediction & daily patterns
- Calendar coverage
The native report format passes Microsoft’s validator. Data checks and visual field references pass independently. Desktop refresh and rendering remain to be verified; the chart above is a Python analysis figure.
Download all DAX measures ↓Download the Power BI runbook ↓
THE METHOD
From source to answer
One row per observed held-out hour, calendar and hour dimensions, 17 DAX measures, weighted error calculations, coverage denominators, and monthly/time-of-day diagnostics.
The Python source exposes this study’s transformations. The complete project download includes shared preparation and evaluation routines.
THE NEXT DECISION
What follows from the finding
Prioritize underprediction diagnostics by time of day, check whether missing hours cluster, and obtain service costs and current operational data before changing planning policy.
Where the conclusion stops
This report reuses the original forecast holdout; it is a diagnostic extension, not fresh validation. Prediction gaps do not measure unmet rental demand or required inventory. Native Power BI refresh and rendering remain unverified because Desktop is not installed.
Related studies may reuse observations or holdouts. These are historical analyses; associations and backtests do not demonstrate commercial impact. Further model tuning needs new, untouched evaluation data.