‹ Back

How this works

The site answers a level question: why does this suburb cost what it costs. That is a different question from what most rent models answer, and a harder one to answer honestly.

The model

For each suburb i and quarter t:

log(rentit) = α + Xiβ + δt + εit

X_i holds the 37 suburb measurements, standardised and sign-aligned so that higher always means more of the thing named. δ_t is a quarter fixed effect that absorbs the citywide time path, so β is a pure cross-sectional gradient rather than a mix of "where" and "when". There are deliberately no suburb fixed effects: they would absorb every static feature and leave nothing to explain. There is no rent history of any kind: no lags, no spatial lag, no asking rents. Houses and flats are fitted separately.

β is estimated by ridge regression on the measurements only; the constant and the quarter effects are never shrunk. The penalty is fixed in advance rather than re-tuned each time the site is rebuilt, so a data update moves the numbers without quietly moving the model underneath them. Every group is in the model at all times; nothing is dropped for being small.

What the dollar figure on each group means

The 9 groups overlap heavily. Almost everything pleasant about a suburb is correlated with being close to the city, so asking "what does this group add once the others are already counted" has no stable answer, because whichever group is counted last gets nothing, and the order is arbitrary. Measured here: Everyday amenities reads +0.64 against its own measurements when fitted alone, and turns negative once location is counted first. Nothing about the group changed. Only what it was counted after.

So the figure shown is a Shapley value: the group's contribution averaged over every order in which the 9 groups could be counted, across all 512 combinations, computed exactly rather than sampled. Where two groups share credit, the credit is split between them rather than handed to whichever the fit happened to favour. The 9 figures on a page add up, by construction, to the whole of what the model says about that suburb.

This is a fair allocation, not a causal claim. It says how the model divides what it can explain. It does not say that changing one of these would change the rent.

Why groups rather than individual measurements

Static features have no within-suburb variation, so they are identified off 103 suburbs, not off the tens of thousands of rows in the panel. Reporting 37 separate figures on 103 observations would be reporting noise: the measurements inside a group move together too closely to be told apart at that sample size. Grouping them is an admission of that limit, not a presentation choice. It is also why each group lists what it contains in rank order and declines to put a number on the parts.

All 9 groups are shown on every page. Socio-economic status was deliberately left out, and the reason is below.

How well it fits

Scored by leave-one-suburb-out cross-validation: fit on 103 suburbs (houses) or 96 (flats), predict the one held out, which is exactly the operation this site performs when you open a page.

	R²	Typical error	Groups with a dollar figure
Houses	0.942	$29/wk	9: Location & access, Dwelling & tenure, Civic & recreation, Safety, Recent arrivals, Education, Coast & environment, Everyday amenities, Development & growth
Flats	0.876	$28/wk	8: Location & access, Dwelling & tenure, Civic & recreation, Recent arrivals, Safety, Education, Development & growth, Everyday amenities

A typical error of $29 a week on houses is not small. The model explains most of the variation between suburbs and misses a lot within them.

Some of that fit is knowing what the neighbours cost

Leaving one suburb out still leaves all of its neighbours in, and suburbs near each other are alike, so the model can partly fill the gap by interpolating. To see how much of the fit that is worth, we held out whole regions instead.

Held out	Houses	Flats
One suburb	0.942	0.876
A whole region (19 of them)	0.930	0.869
A quarter of the city at a time	0.914	0.830
Same-sized groups, picked at random	0.941	0.875

The last row is the control. Holding out groups costs accuracy partly just because the model then trains on fewer suburbs, but scattering the same-sized groups at random costs almost nothing (0.001), so nearly all of the 0.012 drop on houses is geography rather than sample size. The figure at the top of this page is the leave-one-suburb-out one, which is the right question for this site, since when you open a page, the other suburbs really are in the model. But roughly 1.1 points of it comes from the city around a suburb rather than from the suburb's own measurements.

Left-over regional clustering is weak: Moran’s I on the residuals is +0.14 (p = 0.02) for houses. So the model is not leaving obvious regional patterns behind; the dependence is in the measurements themselves, which vary smoothly across the city, not in what the model failed to explain.

An earlier version of this page was wrong

Until recently this site collapsed each group into a single index using principal components, and reported a fit of 0.60 on houses. Checking it against the most obvious possible baseline (distance from the CBD, one number) showed the baseline winning, at 0.82. The collapse was discarding most of the signal: principal components summarise what a group's measurements have in common, which is not the same as what predicts rent.

Rebuilding it so that groups are selected by the model and represented by their own measurements took the fit to 0.942, cutting typical error by about 45% on houses and 25% on flats.

A later correction mattered more than the rebuild. Recorded offences were being split between suburbs using the wrong direction of a postcode-to-suburb crosswalk, which handed large, thinly populated areas a whole postcode's crime: one semi-rural suburb carried 815 offences per 1,000 residents against the 20 actually recorded there. Because that error tracked suburb area, and area tracks distance from the city, it distorted which groups the model could separate at all. With it fixed, 9 of the 9 carry a dollar figure on house pages and 8 on flat pages; before it, three groups could not be priced on houses at all. Residual spread roughly halved, so the “priced oddly?” figures are far less extreme than they were; the largest one on the old site was an artefact of this bug.

When a group gets a number, and when it gets a range

Two questions have to be answered, and they are different. First: can the model measure this group across Adelaide at all? Second: is this suburb far enough from the Adelaide average on it to say which way it goes? A group can pass the first and fail the second, which is why the same group is priced on one page and ranged on another.

Across Adelaide, all four of these must hold, and the thresholds were fixed before any result was looked at:

Reproducible	at least 95% of bootstrap resamples order the suburbs the same way
Not zero	the 95% interval excludes zero
Precise	that interval is narrower than 1.5× the estimate itself
Right way round	the fitted effect points the same way as the measurements the group is named after, tested on the group fitted alone, because testing it inside the full model measured the other groups rather than this one

On this suburb, two more: the direction has to hold in at least 95% of resamples here, and the group’s effect on this suburb must not come out more than three times the citywide spread of that effect. Past that, the figure is larger than anything the model produces elsewhere, and we would rather withhold it than print a number we cannot stand behind.

And a third, which asks the different question of whether this suburb has peers: how far it sits from the other suburbs on that group’s own measurements. Past the standard threshold for an influential point, any figure would be the model reaching beyond the range it was fitted on, and the page says there are no comparable suburbs instead.

The second and third tests sound alike and are not. One asks whether the answer is implausibly large, the other whether the question can be asked here at all, and they disagree more often than they agree: of the suburb-and-group pairs either one stops, only about a sixth are stopped by both. Until recently only the first was checked, while the page reported it in the words of the second.

What a page shows then depends on the suburb's own 95% interval:

A single number	everything above holds
A range	the interval is clear of zero but too wide to pin to one figure
"About average"	the interval contains zero, and usually that is the answer, not a failure: contributions are measured against the Adelaide average and most suburbs genuinely are near it on most groups
Nothing	the interval is wider than the spread of rents across Adelaide, so it would constrain nothing

Where a band contains zero but sits mostly on one side, the page says "probably lower" or "probably higher" rather than "about average", because a band running from −$26 to +$1 a week is not describing a typical suburb.

Collinearity

The groups are correlated with each other: inner suburbs have better transit and more cafes and more schools. Variance inflation factors say how badly that muddies each estimate.

Group	VIF	R² on the others
Location & access	8.50	0.88
Everyday amenities	4.14	0.76
Dwelling & tenure	3.65	0.73
Education	3.62	0.72
Civic & recreation	3.31	0.70
Safety	2.66	0.62
Coast & environment	2.53	0.61
Socioeconomic	1.94	0.48

Location & access at 8.5 is the one to be careful with: it is 88% predictable from the other groups, so some of what it is credited with is shared with them. Splitting credit by Shapley value is what keeps that overlap from being handed to it wholesale, but it does not make the overlap go away.

What was deliberately left out

Socio-economic status. Affluence is downstream of rent about as much as it is upstream, since expensive suburbs become affluent suburbs. A model that "explains" rent with affluence is mostly relabelling rent, so it is left out. Adding it back would improve the fit by 0.004 R² on houses.

School ICSEA scores. ICSEA correlates +0.89 with the SEIFA education index, an affluence measure wearing a school uniform. With SEIFA already out, keeping ICSEA would let affluence back in through the side door. Adding it back is worth 0.004 R².

Both figures used to be far larger (0.28 and 0.10) under the weaker specification described above. Affluence was doing work the other groups could not, because those groups were badly represented. Measured properly, excluding affluence costs 0.4 points of fit on houses and 0.5 on flats. The case for leaving it out got cheaper once the rest of the model worked.

Leaving it out did not remove it

A variable left out of a model does not take its effect with it. Whatever stays behind that correlates with affluence inherits part of it. So we measured how much: refit the same model with the affluence measures added back, and see how far each group’s effect shrinks.

Group	Houses	Flats
Dwelling & tenure	40%	49%
Recent arrivals	40%	30%
Location & access	31%	21%
Civic & recreation	27%	21%
Education	24%	40%
Everyday amenities	16%	20%
Safety	12%	46%
Development & growth	11%	8%
Coast & environment	-6%	-76%

Between 11% and 40% of each group’s effect on houses is shared with affluence, most of all Dwelling & tenure. A negative figure means the reverse: the group’s effect gets larger once affluence is in, because affluence was masking it. This is also why leaving affluence out costs so little fit: the other groups are already carrying much of it. The dollar figures on a suburb page are therefore not affluence-free, and we would rather say so than imply a cleanliness the model does not have. What the exclusion does buy is that no group is named after affluence, so nothing on the page invites you to read “richer suburbs cost more” as if it were an explanation.

Measured on the shipped specification, in sample, as the spread of each group’s fitted contribution across suburbs. The ordering is the finding; treat the percentages as indicative.

Bedroom mix, for flats only. Outer flats are systematically larger, which was showing up as a location effect. Flats are fitted with bedroom-share controls; houses are not, because for houses bedroom mix is itself a location proxy, since the share of three-bedroom-plus houses correlates +0.52 with distance from the city.

The forecast

A separate model, on a different target: the quarter-on-quarter change in log rent, not the level. Prediction intervals come from split conformal prediction, calibrated by rolling origin. For each quarter, the model is fitted on data ending two quarters earlier, early-stopped on the quarter before, and calibrated on the quarter itself, so no residual has ever seen the row it describes.

On two quarters held out of the calibration entirely, 80% intervals covered 81% of houses and 85% of flats; 90% intervals covered 91% and 93%. These are development-window figures. The frozen test window was spent once, on an earlier specification of the forecast, and is not reused, so these numbers carry less weight than a one-shot test result would, and we would rather say so than quote the old model's test figures next to this model's forecasts.

Calibrating before the 2021 market break and applying afterwards fails badly, which is why calibration uses post-break data only. Flats carry much wider intervals than houses: an 80% interval spans 18% of the rent on flats against 8% on houses, and at 90% it is 2.2 times wider, because bond volumes are thinner and the series is noisier.

The interval is the product here, not the point estimate. See what this can’t tell you for why.

Where the rest of it is

Every model, diagnostic and failed specification behind this site, including the ones that overturned earlier conclusions, sits in the research repository alongside the changelog that orders them, which is not public yet.