Do Jobs & Approvals Predict AU Property Prices? — Four Regional Signals, Backtested

Four families of regional economic signals, each from a different ABS dataset and each measured at a different geographic granularity. The underlying question was the same in every case: does this regional number, known on the day a suburb is being scored, help rank which postcode within Australia will outperform the broader market over the next two years?

The panel ran from August 2009 to April 2024 across 2,271 Australian postcodes. After joining each regional signal to its constituent postcodes by month, the dataset comprised 310,575 postcode-month observations. Forward returns were computed as the change in median price between the scoring date and 24 months later.

Every signal was scored using within-date Spearman information coefficient. That phrase carries the entire methodology, so it is worth unpacking. On each of 177 scoring dates, every postcode received a value for the signal and a forward two-year return. Spearman rank correlation was computed between those two columns — meaning we measured whether postcodes ranked highly on the signal also ranked highly on subsequent returns, regardless of magnitude. The coefficients across all dates were then averaged.

Two design choices matter here. First, within-date: the correlation is computed across postcodes on the same date, not across time. This neutralises the macro tide — a signal that merely tracks the national property cycle gets no credit. Second, market-cancelled: the per-date median forward return is subtracted before correlation, so what we are measuring is relative outperformance, not absolute price growth. This is what an investor actually cares about when picking between suburbs.

A perfect signal would score +1.00. A useless signal averages 0.00. Anything between roughly ±0.05 is statistical noise in a panel this size. To survive into the suburb-scoring formula, a signal needs both a meaningful magnitude andstability across eras — pre-2015 versus post-2020.

The table below shows the mean information coefficient (IC) for each signal, the t-statistic measuring its statistical significance, the number of scoring dates available, and the IC split by era. The Verdict column is what the suburb-scoring pipeline does with each signal next.

Panel: 310,575 postcode-months across 2,271 postcodes, 177 scoring dates between August 2009 and April 2024. IC = within-date Spearman rank correlation between signal and forward two-year market-cancelled return.

Read across the table and the picture is consistent. Every mean IC sits within ±0.10 — the strongest in the set is regional unemployment at −0.09, and even that signal looks unstable once you split it by era. Statistically significant t-statistics here are a function of sample size, not effect size: with 148 scoring dates and tens of thousands of postcode-months per date, a coefficient of 0.05 will register as “significant” while remaining economically useless for picking suburbs.

The headline finding sits in the era split. Regional unemployment had a within-date IC of −0.23 before 2015. That is a meaningful number — high-unemployment SA4s genuinely underperformed their peers in the post-GFC decade, and the sign pointed where intuition expected. After 2020, the same signal flipped to +0.07. High-unemployment regions became modest outperformers. Participation rate did the mirror image: +0.21 pre-2015 to −0.07 post-2020.

A directional reversal of that magnitude is not a calibration problem you fix with a different lookback window. It indicates the signal was tracking something else entirely — something that itself changed. Three candidate explanations are consistent with the data, and likely all three contributed:

Industry-mix HHI shows the same pattern in milder form. Pre-2015 the 12-month change in concentration had essentially no signal (−0.0035). Post-2020 it jumped to +0.12. Whether that holds in the next regime is unknowable from this data. What we can say is that a signal which only appeared after 2020 is a candidate for further investigation, not a load-bearing input to the scoring formula.

Even setting era stability aside, regional macro signals face a structural problem when used for suburb selection. SA2, SA4 and LGA boundaries each contain many postcodes — a typical SA4 spans dozens of distinct suburbs — and every postcode inside a region inherits the same regional reading.

That collapses the within-region rank correlation to zero by construction. If forty postcodes in the same SA4 all share the same unemployment value, the signal cannot distinguish which of those forty will outperform. Suburb selection — the question an investor is actually answering — happens within regions, not between them. A signal that only ranks regions cannot rank suburbs inside a region.

This is why the within-date IC is the right yardstick. It measures the signal’s ability to rank postcodes against each other on the day a portfolio decision is being made. Regional macro inputs fail this test because their resolution does not reach below the regional boundary.

Suburb-level outperformance, in our backtest, was best predicted by two suburb-level inputs: affordability headroom (the suburb’s median price relative to its capital city median) and current boom detection (a composite of price momentum, days on market, vacancy rate and growth strength). Both are measured at the suburb — not regional — level, and both held up across the same pre-2015 / post-2020 split that sank the labour signals.

The fuller treatment of those two surviving signals is in metrics that predict suburb booms and the broader question of why expert frameworks disagree is covered in why property experts disagree. The short version: signals that survive era splits and resolve at the suburb level are rare. Most of what gets cited in property commentary fails one or both tests.

You can look at the same ABS releases we do and reach the opposite insight. The labour figures, the approvals data, the industry-mix tables — they are all public. They have all been cited in property commentary for decades. The reason the conclusions diverge is not access to data; it is what you do with it.

The methodology is what separates a signal from a coincidence. Within-date validation strips out the macro tide. Market-cancellation asks the right question (relative, not absolute, returns). Era-splitting catches signals that only worked in one regime. A backtest that does none of those things will surface false signals consistently — and an analyst who ignores the era split will confidently recommend a labour-market framework that has been inverted for five years.

For practical suburb selection, the implication is narrow: regional unemployment, regional participation, regional approvals and local industry concentration should not be primary inputs. They might be useful as context, or as risk overlays at the city level. They are not the place to look for an edge in picking the next outperforming postcode.