IPMVP Option C in practice: what a defensible savings claim actually requires

Option C — whole-facility measurement — is the most used and most abused of the IPMVP options. It is attractive because the meter is already there: you have the utility bill or the main incomer, a before period and an after period, and the difference looks like savings.

It is not. The difference is the difference. Savings is what remains after you have explained everything that changed for reasons other than the thing you did.

What follows walks the claim through with the arithmetic in view. The numbers are illustrative, not a customer result, but internally consistent — check each step.

The structure of the claim

An Option C claim is one equation:

Savings = (Baseline-period energy, adjusted to reporting-period conditions) − (Reporting-period energy)

Every argument about M&V is an argument about the word adjusted. The protocol is maintained by the Efficiency Valuation Organization, whose Core Concepts 2022 edition marks "the 25th anniversary" of the protocol. Those years have not made adjusted easier; they have made it better documented.

1. Choose the option before you choose the maths

People reach for Option C because whole-building data is the data they have. That is a reason, not a decision. IPMVP is organised by where you draw the line: EVO groups the options as retrofit isolation (A and B), whole facility (C) and calibrated simulation (D).

The decision rule is signal-to-noise. Option C measures every source of variation in the building, not only yours. If your measure moves 3% of consumption and unexplained week-to-week variation is larger than that, no regression will find it; the response is a tighter boundary, not a better model. Tightening has a price — you must then account for interactive effects, because a better chiller changes the pumps and fans around it.

The boundary is an energy balance, not a diagram

The test of a boundary is an energy balance: everything entering it is either measured, or explicitly out of scope and stated. A site with a main incomer, a standby generator, a solar array behind the meter and a district chilled-water connection has four energy streams and usually one of them in the model. Generator hours sit in a logbook. Solar is netted invisibly at the incomer, so a sunny year reads as a savings year. District cooling is billed in ringgit, not kWh, and never reaches the analyst.

The remedy is dull and effective: list the streams, mark each measured, stipulated or excluded, reconcile against the meters you have. An exclusion is inherited by the claim and has to be stated — the same discipline as putting coverage alongside every aggregate.

2. The baseline period has to be defensible on its own

• Long enough to span the driver range. Twelve months covers a full weather cycle. A baseline built on a cool six months under-predicts a hot reporting period and your "savings" go negative through no fault of the measure. The open-source CalTRACK methods make this a data-sufficiency rule: "Consumption and temperature data should be sufficient to allow for a 365-day baseline period."
• Representative of normal operation. A baseline containing a three-week shutdown is two regimes averaged into one that describes neither.
• Declared granularity. Monthly bills give twelve points; hourly intervals give 8,760, and support a far stronger model.
• Complete, or explicitly incomplete. CalTRACK caps missing data at "37 days (10%) for billing and daily methods" — because analysts otherwise model whatever survived and call it a year.

The n you actually have

With monthly bills and two drivers you are fitting three parameters to twelve observations. That model will look excellent — high R², small residuals — and generalise badly, because it has barely enough degrees of freedom to be wrong in. Berkeley Lab's accuracy assessment of automated M&V techniques found "median coefficient of variation of the root mean squared error (CV(RMSE)) was less than 25% for every model tested when twelve months of training data were used". That is what ordinary interval data supports across a population of buildings — a different regime from twelve bills.

3. Driver selection is where the honesty lives

An independent variable earns its place by two tests: a physical mechanism, and statistical significance. Both, not either. Outdoor dry-bulb temperature almost always qualifies for a cooling-dominated facility, usually as cooling degree days at a fitted base. Humidity matters more than people expect in the tropics: a chiller dehumidifying outdoor air at 80% relative humidity is doing work dry-bulb does not describe.

What does not qualify, however good the correlation:

• Indoor temperature. An outcome of how the plant ran, not a driver of it. Include it and you are partly regressing energy on itself.
• Anything downstream of the measure. If the ECM changes the variable, the variable cannot normalise the ECM. Chilled-water supply temperature is the usual offender.
• Significance without mechanism. With enough candidates something correlates. That is what p-hacking is.
• A driver you cannot measure in the reporting period. A model you cannot feed is not a model.

4. The claim, worked end to end

An air-conditioned commercial block on hourly interval metering. The measure is chiller plant optimisation — sequencing, condenser-water reset, variable-speed primary pumps — commissioned July 2025.

Baseline period and fitted model

Baseline: 1 July 2024 to 30 June 2025, 365 complete days, fitted at daily resolution. Two drivers: cooling degree days at a fitted base of 24 °C, and an occupied-day indicator.

kWh/day = 4,180 + 96.5 × CDD₂₄(day) + 2,640 × occupied(day)

Read it physically before you read it statistically. The intercept is the building with nothing to cool and nobody in it; the occupied-day term is the cost of opening the doors; the degree-day term is the plant's response to weather. All three have magnitudes a facilities manager can argue about, which is the test a coefficient should pass before anyone reports a number from it.

Fit on daily data: adjusted R² 0.86, CV(RMSE) 6.8%, NMBE +0.1%. Mean daily consumption is 6,316 kWh, so 6.8% is a prediction error of about 429 kWh per day. Hold on to that figure; it decides whether the claim survives.

Reporting period and the adjusted baseline

Reporting: 1 August 2025 to 31 July 2026. The year was hotter — 3.90 CDD₂₄ per day against 3.40 — with 252 occupied days. Feed those conditions to the baseline model and it says what the building would have used had nothing been done.

Simple subtraction of one year from the other gives 2,305,457 − 2,142,000 = 163,457 kWh, or 7.1%. The naive figure understates the real result by roughly an eighth, because the reporting year was hotter and the building paid for the weather out of its own savings. The adjustment corrected the claim upwards — the direction people forget is possible. We work through that asymmetry in Last year's kWh is not your baseline.

5. Uncertainty, and why a small claim is not a claim

186,396 kWh is a point estimate from a model wrong by about 429 kWh on a typical day. The savings figure is a sum of 365 daily predictions; if the residuals were independent, the standard error of that sum would be the daily RMSE scaled by the square root of the number of days:

SE = 429 × √365 = 8,196 kWh

A 95% interval is roughly ±1.96 standard errors, so:

Savings = 186,400 kWh ± 16,100 kWh (95%), or 8.0% ± 0.7 percentage points.

The ratio of band to savings — about 8.7% here — is what M&V practice calls fractional savings uncertainty, and it decides whether a claim is reportable. EVO runs a statistics and uncertainty committee and publishes Uncertainty Assessment for IPMVP (EVO 10100:1:2019); ASHRAE's Guideline 14-2023 carries "an expanded discussion of uncertainty calculations". Uncertainty is half the answer, not a footnote.

Where the simple calculation is optimistic

That formula assumes independent daily errors. They are not: whatever the model is missing — a stuck damper, a humid week, plant left in manual — persists for days, so the effective sample is smaller than the calendar count and the true interval wider. Berkeley Lab's advanced M&V research reports that "typical methods for estimating savings uncertainty over the post-installation period tend to underestimate the uncertainty, and that this tendency is stronger for hourly models than for daily models". At a third of the nominal sample the interval here widens to roughly ±27,900 kWh — still clear of zero, which is all this claim needed.

The bottom row is the one to dwell on. The point estimate is 46,600 kWh; the interval runs from about −700 to about 93,900 kWh. "46,600 kWh saved" out of that model is the midpoint of a range that includes the building using more energy than before. The correct output is a refusal with a reason — too small to verify at this boundary with this data. That is what savings verification means done properly, and the principle behind refusing to render a missing reading as zero.

6. Non-routine adjustments are the audit's first stop

A non-routine event is any change unrelated to the ECM that affects energy: a floor let to a new tenant, a server room added, a line retired, a shift pattern changed. Each requires an explicit, documented adjustment. EVO's application guide on non-routine events and adjustments (October 2020) calls them "a key risk element for energy efficiency projects using meter-based M&V methods" — a risk that sharpens "when M&V is the basis for monetary transactions".

Say a tenant commissions a data room on 1 February 2026, inside the boundary, sub-metered at 41,000 kWh to year end. The baseline model never saw that load, so its counterfactual is the building without it while the meter includes it — charging 41,000 kWh of unrelated load against the chiller project.

The adjustment raises the claim by 22%. It is legitimate only because it is derived from a measurement rather than an estimate of what a data room that size probably draws. Remove the sub-meter and the adjustment becomes an assertion; the defensible action is then to report the unadjusted 186,400 kWh and state that it is understated by an unquantified amount. A conservative claim you can defend beats an accurate one you cannot.

Two failure modes, one in each direction. Silent absorption: a tenant moves out, consumption drops, the drop is claimed as savings — usually not deliberate, because nobody told the analyst. Unbounded adjustment: an adjustment with no stated basis is a free parameter that produces whatever answer its author wants.

The control is a timestamped log of facility changes, kept by operations rather than the analyst and reconciled against the residual plot. A residual step with no entry in the log is an unexplained change and should be reported as one. In buildings whose BMS you did not commission, that log is often the only record that anything happened.

7. Fit statistics, and what they do not tell you

CV(RMSE) and NMBE are the statistics everyone quotes, and ceilings on them are commonly applied in programme rules — which ceiling depends on the document your programme adopted, the data interval and the modelling approach. Be careful quoting a number without its source: the criteria applied to calibrated simulation are not those applied to whole-building regression, and the two are conflated constantly. Whichever you are held to, the statistics are necessary and nowhere near sufficient.

The check none of these perform: plot the residuals against time and against each driver. If they have shape — a seasonal wave, a step, a fan widening with load — the model is missing a term, and no summary statistic will say so. A model can hit every threshold it is held to and still be wrong in precisely the period you are claiming savings in.

8. What to do when the data is missing

Meters fail, gateways drop, a week of the baseline is absent. Three honest options: exclude the period and state its effect on n; fill it from a genuinely independent measurement and label it a fill; or refuse to report for that period.

Interpolation is not on the list. Neither is a zero. A zero in an energy series claims the building drew no power, and unless it was de-energised the claim is false. CalTRACK encodes this: "Values of 0 are considered missing for electricity data, but not gas data", and in the reporting period "counterfactual usage is not calculated when daily temperature data is missing". A method that refuses to compute is doing its job.

9. What an assurance provider actually asks for

When the claim leaves your organisation — to a lender, an ESCO counterparty, an incentive programme, an assurance engagement — someone asks for the working. The US Federal Energy Management Program, whose M&V Guidelines version 5.0 (October 2024) cover "verifying and quantifying the savings resulting from energy-efficient equipment", is blunt about where the burden sits: FEMP calls the M&V plan "the single most important item in an energy savings 'guarantee'". The request list is short and unforgiving:

1. The M&V plan, dated before installation, naming the option and the boundary.
2. Raw interval data for both periods, exportable, with meter identifiers.
3. The model: form, coefficients, fit statistics, residual plot.
4. Driver data with provenance — which weather station, which occupancy source.
5. Every non-routine adjustment, with the measurement it came from.
6. Every excluded or filled interval, with its reason and effect on coverage.
7. The uncertainty, at a stated confidence level.
8. Evidence the baseline has not been edited since certification.

Item 2 separates platforms. Without the underlying intervals, the rest is a report about a report. A customer should be able to export every interval behind every figure and reconcile it against their own source — the only version of "trust us" that survives an auditor.

10. How claims fail, and the tell for each

Each tell takes minutes to check, and months to defend once missed. That asymmetry is why the discipline is worth the effort up front.

Why this matters beyond the energy bill

A verified saving is also a carbon number. The GHG Protocol Scope 2 Guidance governs accounting for "emissions from purchased or acquired electricity, steam, heat and cooling (called 'scope 2 emissions')", and a reduction claimed in a sustainability report is the same kWh multiplied by a grid factor. If the kWh is an assertion, so is the tonne of CO₂e — and the assurance applied to the disclosure is usually stricter than the review applied to the energy report. That is the argument in ESG reporting is a cost centre, and behind how we structure reporting-grade measurement.

The short version

A defensible Option C claim states: the option and why it beat A, B and D; the boundary as an energy balance, every stream marked measured, stipulated or excluded; the baseline period and why it is representative; the drivers and their mechanism; the model form, coefficients and fit statistics including a residual plot; every non-routine adjustment with the measurement behind it; every excluded or filled interval with its effect on coverage; and the savings with an uncertainty range at a stated confidence level.

If a report gives you a single number and a percentage, it has not made a claim. It has made an assertion, and the difference will surface the first time someone with an incentive checks.