Why we deleted our ten year forecast

An early build of this tool projected a purchase ten years ahead. You gave it a price, it applied a growth rate, and it printed what the property would be worth in 2036. It was the most convincing screen we ever built, and we removed it.

The problem was not bad maths. The problem was that the answer was almost entirely set by an assumption the reader had never made. Move the rate from 5% to 7%, both easy to defend, and a ten year figure moves by more than a fifth. Every seller in this industry knows which of those two numbers to pick. A page that turns a rate nobody argued about into a big number is not analysis. It is a brochure with a calculator in it.

That reasoning has not changed. It is why the forecast here runs one year ahead, not ten. The further out a property forecast goes, the more of it is assumption and the less of it is data. Past about a year, a suburb-level forecast is really a statement about interest rates and migration policy with a suburb's name on it.

How the forecast is built, part by part

The twelve-month price forecast is a sum, not a black box:

state baseline  +  region tilt  +  momentum tilt  +  yield tilt  +  reversion tilt  +  demand tilt

Six numbers add up to the answer, and every one of them is published as its own figure on the market's page. The baseline is what the model expects for that state and property type as a whole. The tilts are the adjustments that move it: how the region around the suburb has been going, how the suburb's own trend looks, whether its yield (rent as a share of price) is high or squeezed compared with similar suburbs, whether its price has run far above its own long-run path, and how hard buyers are competing for what is listed.

This is the part that matters, and it is something a single predicted number can never do. Say a market is forecast at +9% and eight of those points come from momentum. That is a completely different bet from one where those points come from demand and yield. The second kind is the one that tends to hold up when next year arrives. Because the parts are separate, you can decide that this market's momentum came from two big sales, take it out, and still use the other five.

The rent forecast is built the same way. If you plan to hold rather than flip, it is the more useful of the two. Rent is the half of the return that arrives whether or not the price rises.

Rule one: every forecast comes with its range

Every forecast on this site has a low and a high, plus the odds that growth is positive at all. The range is not decoration, and it is not a disclaimer. On many markets it is wider than the forecast itself, and that is the single most useful thing it can tell you.

Two suburbs both forecast at +8% are not the same bet if one has a band of +5% to +11% and the other runs from −4% to +20%. On the second, the honest reading is "this model does not know", and the width of the band says so. That is why every market page draws the range before the number.

Rule two: we tell you where every forecast comes from

Not every suburb has enough price history of its own to fit a model to. Where it does not, the forecast leans on the region and the state. That is a fair thing to do and a dishonest thing to hide, because the result is a regional expectation printed next to one suburb's name.

So every market publishes an own-data weight: the share of its forecast that came from the suburb itself rather than from the area around it. Read that figure first in that section. It decides how literally to take everything under it. Markets where the model leaned mostly on the region are marked low confidence, and the plain-English summaries on the cards refuse to put their forecast into words at all. A sentence sounds certain in a way that a number with a band does not.

See the region behind every suburb’s forecast

One of the six parts above is a region tilt. Until recently you had to take that on trust: the forecast said the area moved it, and the area was nowhere to be seen. Now every market shows its region measured the same way the suburb is: the same growth windows, the same rent series, the same forecast, the same backtest (a re-run of the model on a past year to see how it did), worked out across every suburb that shares the name.

That does two things. It makes a suburb figure something you can judge: up 14% in an area that rose 12% is two points of noise, and up 14% in an area that rose 2% is either a real find or nine sales in one good street. And it shows how widely the area moved: the share of its suburbs that rose at all, the typical suburb in it, and the gap between the quarter mark and the three-quarter mark. A region can post 8% because forty suburbs each rose 8%, or because two of them doubled. Those are opposite investments wearing the same number.

The part that matters most for a forecast is the last one: how accurate this model has been in that specific region. A national error rate tells you what the model is worth in general. The hit rate, and the rate at which it beats a naive guess, on the suburbs around the one you are looking at, tell you what it is worth here. Where that beat rate sits near or below half, the page says so in plain English instead of letting the forecast suggest otherwise.

The region makes one more thing possible. A thin suburb’s own growth figure is mostly noise, and its region’s figure is mostly a fact about somewhere else. So every market also publishes a blended figure: the two pulled together, based on how much that market’s own record has actually earned, given how many sales sit behind it. On a deep market it is the suburb’s own number. On a thin one it is close to the region’s. That is what a careful analyst does in their head, and it is what stops a ranking being topped by nine sales in a good street.

Rule three: we publish our own error rate

We run the same model again on data that ended twelve months before the data it runs on now. Then we compare its predictions with what actually happened. Those figures are on every market page.

Across the dataset the typical miss is about 6.7 percentage points, and the model gets the direction right about four times in five. In the past it has also been more gloomy than the market turned out to be, by around four points on average.

Those are not flattering numbers, and that is the point of publishing them. A forecast without an error bar is a guess with a decimal point on it. Anyone can print +12% next to a suburb name. What tells you whether to act on it is knowing that a figure from this model has been off by six or seven points about half the time, and that its real value is in ranking markets against each other, not in the second decimal place of any one of them.

What we will never do with a forecast

• No ten year projection. Not in the free view, not in the paid data, not in the export. The three-year figure fades back toward the long-run baseline instead of compounding the twelve-month rate. Compounding a one-year forecast for three years is how a modest number turns into a spectacular one without gaining any new information.
• No forecast dressed up as a fact. Past figures and forecast figures sit in separate groups, on separate panels, with different wording. Nothing on this site prints a prediction in the same row as a measurement.
• No unfiltered forecast ranking. Sort every market in the country by a prediction and the top of the list fills up with the ones the model knows least about. Thin history makes wide bands, and wide bands make flattering midpoints. The forecast leaderboards only include markets with real history behind them, and they say so.
• No forecast in the dollar figures. The gain calculations still apply growth that has already happened to your own purchase price. That number cannot be wrong, and it has no assumption in it for anyone to bend.

How to use the forecast to find stronger suburbs

Rank on forecast return to see where the model thinks the next year is. Then read the three columns next to it before you shortlist anything: the range, the odds, and the confidence. A market that passes all three is worth a look. A spectacular forecast with a wide band and a low own-data weight is the model telling you it is guessing, as politely as it can.

Then check it against the record, which is on the same page. The forecast is drawn from the history. Where the two disagree, the disagreement is the finding. A market that grew 15% last year and is forecast to do nothing next year is telling you something no single figure on any property portal will.