Skip to main content
Forecasting · Diagnosis

Know what your buildings will use. And why they didn’t.

Forecaist forecasts your consumption meter by meter, then explains the gap when reality disagrees. Caught early, a deviation is a setting to correct. Caught in month six, it is a bill.

Book a demoHow we measure

Built on the metrics energy buyers already use: ASHRAE Guideline 14 and IPMVP.

Forecast vs actualone meter, one week, illustrative
ActualForecast

What we actually do

Three things, in the order they depend on each other. A forecast nobody trusts is not worth diagnosing, and a deviation nobody can price is not worth chasing.

01

Consumption forecasting

Know what the building will use, before it uses it.

From your consumption history, the weather, and the calendar — working days, weekends, public holidays — we build a model that predicts your coming consumption. Every model is tested the same way: measured against a naive baseline — “this hour will be like the same hour last week” — then validated on data it has never seen, so the score you are shown is the score it earned on unseen data rather than on the data it was fitted to.

  • One model per meter, not one per building
  • Scored against a naive baseline every time
  • Validated on held-out data, read once
02

Deviation detection and diagnosis

When reality and forecast diverge, find out why.

As soon as a significant gap opens between actual and predicted consumption, the system flags it. Identifying the probable cause — unusual weather, an unusual day, a change in how the building is used, drift in the heating or cooling system, or a fault in the meter itself — is the half we are building now, and we will tell you which parts are live before you rely on them. The point is to catch the problem — and therefore the avoidable cost — early, before it compounds over months.

  • Statistical control limits, not a fixed percentage
  • Each finding carries the numbers behind it
  • “Not enough evidence” is a valid answer
03

Caught while the fix is still small

Act while it is still cheap to act.

Identifying drift quickly is what lets you act before the extra cost becomes significant — whether that is a setting to correct, a piece of equipment to check, or a habit to adjust. A fault caught in week one and a fault caught in month six are the same fault at very different prices.

  • Deviations sized in kWh, and in money once tariffs are known
  • Early enough that the fix is still small

Measured the way you would check it

Energy forecasting is easy to make look good. These are the rules we hold ourselves to, and they are the reason our accuracy claims are more modest than some you will be shown.

Every model is scored against a baseline, win or lose

The naive baseline is “this hour will be like the same hour last week”. It is harder to beat than it sounds. We report improvement over it rather than a bare error figure, because an error figure on its own tells you nothing about whether a model was needed.

The test window is read once

We held back three months the models had never seen, froze the configuration, and read it exactly once. A test set you check repeatedly stops being a test set — it quietly becomes something you tuned against.

Time-respecting validation only

Folds run forward in time. Shuffling energy data lets a model learn from next Tuesday to predict last Monday, which flatters every score it produces and none of the forecasts it makes.

Metrics your engineer already recognises

CV(RMSE) and NMBE, the ASHRAE Guideline 14 and IPMVP measures used in measurement and verification — not R², which can look excellent while the model is systematically wrong.

What our models scored on our pilot

Three buildings across two sites, five meters, Budapest. Day-ahead forecasts, scored on the weather forecast that was actually available at the time. Measured over 9 March – 1 June 2026 — twelve weeks the models never saw during development, read once.

5 of 5

meters pass ASHRAE Guideline 14

hourly CV(RMSE) at or under 30%, absolute NMBE at or under 10%

4 of 5

beat the same-hour-last-week baseline

by 22.7% to 44.8%; the fifth lost to it by 3.2%

9.3%

best CV(RMSE) on the held-out window

on a base-load meter; the hardest cooling meter scored 28.7%

Questions we get asked

Tell us about your buildings

Send us a meter export and we will score it against the same baseline and the same acceptance thresholds we used on our own pilot, and show you the working.

We reply ourselves, usually within two working days.

Whichever you pick, nothing needs installing and we source the weather ourselves.

A couple of sentences is plenty. The detailed questions come by email.

Book a call while you are here

Optional. Pick a slot and it is booked straight away — a Google Meet invitation lands in your inbox before you close this page.

Pick a day

Loading available days…

We use this to reply to you and for nothing else. No newsletter, no sequence. How we handle it: Privacy.