How I work out what your home will cost

Every pound figure I put in front of you comes from a model of your home rather than a national average. Here is what goes into one, how I test whether it is any good, and what it measured out at when I checked.

What goes into a model of your home

• Your own meter, half hour by half hour. If your meter is connected, the readings come from the national smart-meter network, so I can see when you use energy and not only how much. That is the difference between knowing your annual usage and knowing your evenings.
• The weather over your postcode, past and forecast. Historic and forecast temperature from Open-Meteo, matched to your coordinates. A cold week is the single largest thing that moves a heating bill, so a forecast that ignores the weather is guessing.
• Your building, from its energy certificate. Where a home has an Energy Performance Certificate, EPC Open Data gives its band, floor area and heating type. Where there is no certificate — most homes have none — I use the certificates on your street instead, and say so.
• The tariffs actually published where you live. Your grid-supply region comes from postcodes.io; the rates come from suppliers' own published prices, the Octopus public API and the Ofgem cap, per region. Not a national headline rate.
• Your own bill, if you show me one. A photo or three typed figures replaces the regional rate with the rate you are genuinely paying. The image is read on our own hardware, held in memory and never written to disk.

That is the whole list. Every third party that receives anything is named on the sub-processors page.

I do not fit one model. I run several and let your home pick.

Homes fail to be predictable in different ways. A flat that uses much the same energy all year needs nothing clever; a poorly insulated house with a heat pump is almost entirely a weather story. So I fit several candidate models to your readings and keep whichever one is actually winning on your home, per fuel.

• Your steady average. Your own daily average, projected forward. Simple, and for a home that uses much the same energy all year it is genuinely hard to beat.
• Your seasonal shape. I learned how your use rises and falls through the year, then price each remaining day at where it sits on your own curve.
• Last year, adjusted. This month last year in your home, rebased on how this year is running so far. Your habits do most of the work.
• A blend. Your seasonal shape and last year's same month, averaged — two views of your home that fail in different ways, which makes the pair steadier than either.
• Your seasonal shape, steadied. Like the seasonal model, but a cold first week is not allowed to speak for the whole month. Early in the month I lean on your long-run pattern and hand over to what is actually happening as the days come in.
• Your heating response. I measured how much more energy your home uses for each degree it drops outside — that slope is your insulation and your heating, and it is specific to your house.
• Your full model. Weather, the time of year, weekdays against weekends, and what you did this week last year, weighed together. This one needs more than a year of your history, which is why it takes a while to earn its place.

The choice is remade every night. If a different model starts winning on your home, your home moves onto it without anybody deciding to.

How I test it: on months the model was not allowed to see

A model graded against the data it was fitted to will always look brilliant, and tells you nothing. So every candidate is scored the other way round: fit it on part of your history, then make it forecast a stretch it has never seen, then compare that forecast with what your meter went on to record. That is the only score that decides which model your home uses.

On top of that, the figures I actually showed you are written down when I show them. Once the month finishes, each one is graded against the real total, so what you can see on your own accuracy page is a record of what I told you, not a re-run with hindsight.

What it measured out at

Six figures rather than one, because the useful thing about forecast accuracy is how it changes as the month fills in. Early in a month I am forecasting weather nobody can forecast. By the end I am mostly adding up what your meter has already sent.

A whole month, forecast on the fifth, is a twenty-five-day weather forecast for one building. Anyone quoting you a single flattering percentage for that is quoting the end of the month and not telling you.

What that means in pounds

A percentage is hard to feel, so here is the same measurement in money. This is not a percentage applied to an average bill: the backtest prices the forecast and what actually happened at the same rate, so the gap between the two is measured in pounds directly. Each row says what the miss is an error on, and how much evidence sits behind it.