Method

Measured, or refused. Never invented.

Every figure is computed from readings that were actually taken. Where a reading is missing, Radr says so and says why. It does not substitute a default, an average or a zero.

Missing data

Is a missing reading the same as zero?

No. Zero kWh is a measurement: the meter reported and saw no use. A missing interval says nothing about the building, only about the instrumentation. Treating the two alike reports use that was never observed, and reports it low, which flatters a saving.

  • A gap stays a gap, in storage, in every aggregate and in the export.
  • Every total carries the share of its window that was measured.
  • A figure that cannot be computed honestly is refused, with the reason.

Radr never interpolates across a gap, zero-fills it, or stamps data of unknown age with the current time.

Daily roll-up · incoming supply
Mon · coverage 100%1,842 kWh
Tue · coverage 100%1,796 kWh
Wed · gateway offline 03:00–09:00, 6 of 24 h missingno figure
Thu · coverage 100%, +8% vs model1,903 kWh
Fri · coverage 100%1,655 kWh
Week to date · 114 of 120 h measured8,472 kWh · 95.0%
Wednesday is absent, not zero. The week total is still published, over the 114 hours that were measured. Sample data, not a customer result.
Hourly coverage · one working weekSample data
measured absent — not estimated, not zero-filled95% coverage · 6 of 120 intervals absent
One cell per hour. The unbroken run is the Wednesday outage: not zero, not estimated.
Data coverageSample data
What was measured, and where data is missing.

Baseline method

How is a baseline chosen and tested?

Drivers: mechanism first, then statistics

A variable enters the model only if a physical mechanism moves consumption through it (weather, occupancy, production) and the relationship is significant over the baseline period. Either test alone is a trap.

Indoor temperature is an outcome, not a driver

It is produced by the system being evaluated. In the model it absorbs the very change being measured, and the saving disappears into the regression. Setpoints, valve positions and plant runtime belong in diagnostics.

Fit criteria

Criteria from ASHRAE Guideline 14: CV(RMSE) of 20% or better on monthly data and 25% or better on hourly, with NMBE within ±5%. Radr also plots residuals against time, each driver and fitted value, because a model can pass the numbers and fail the plots.

Non-routine adjustments

A floor taken out of service, a new tenant, changed hours: changes the baseline period cannot explain. They are adjusted for explicitly, with the evidence attached, or a new baseline is set. Never silently.

Baselines are versioned records

A certified baseline stores its period, drivers, coefficients, fit statistics, intervals and coverage, with its author and date. A change supersedes it with a new version; the old one stays readable. Each change is written to an audit trail.

Evidence you can check

The intervals behind any figure can be exported and recomputed against your own meter reads or invoice.

Baseline fit · daily kWh against cooling degree daysSample data
driverresponse
Observations
42 days
CV(RMSE)
8.90%
NMBE
−0.01%
R²
0.83
Each point is one day of the baseline period. R² alone is not enough: the residuals must have no shape.
Fit evidenceSample data
model error, %-100+10JanFebMarAprMayJunJulAugSepOctNovDec
The intervals behind the fit.

See which figures we would refuse

Send a sample export. We will say which figures are measured and which we would decline to publish. How savings are verified

Send a sample