How Accurate Are Property Forecasts?We Built One. Tested It. Abandoned It.

The distinction that rarely gets made in property commentary, but matters more than almost anything else:

These look like variations on the same question. The data requirements are completely different, and the accuracy gap is enormous. Here’s what the testing showed.

The first formula used the metrics that property commentators reach for automatically: infrastructure spending, population growth, building approvals. The theory was that these leading indicators could identify boom suburbs 12–18 months before the market caught on.

A coin flip. With seven inputs and a lot of extra work.

The core problem: infrastructure projects affect whole corridors over decades, not individual suburbs over 18-month windows. Population growth data moves too slowly and is too coarse to rank suburb A against suburb B. Building approvals are measured at state level — they tell you about supply pressure in a broad area, not whether a specific suburb’s buyers are running out of options. Vendor discount data barely exists at suburb level from free sources.

None of these signals have the resolution needed to rank individual suburbs against each other over an investment-relevant time horizon. They’re meaningful context for understanding broad market direction. They have near-zero predictive power at the suburb level.

After the prediction formula failed, we tried something more sophisticated: a proper suburb forecaster. A model that would predict 3-year forward capital growth for every suburb, with confidence ranges. Six inputs: annual growth rate, how affordable the suburb was relative to its city median, whether a boom had already been detected, 5-year growth momentum, the national market cycle, and the recent acceleration in growth.

We ran it across thousands of suburb-months of historical data. The headline result was compelling.

For a moment, this looked like a working forecaster. Then we ran the test that changes everything.

Most property forecasts are evaluated by pooling all their predictions together and measuring how well they correlated with outcomes overall. That feels rigorous. It’s not.

Here’s the problem: when you’re deciding where to buy, you’re making that decision in a specific month. You are not choosing between a suburb available in 2017 and one available in 2023. You’re choosing between two suburbs that are both available to you right now. The question the model needs to answer is: within this month, which suburb will outperform?

So we split the analysis within each individual time period and asked: does the model’s ranking of suburbs predict their relative performance in that same period?

Negative 0.08. Worse than random.

The pooled 0.42 was a statistical illusion. Boom years produced high growth broadly — the market tide lifted every suburb. Our model happened to give higher scores in boom years. So the pooled correlation looked strong. But within any given month, the model had no ability to tell you which suburb would beat the others. It was ranking time periods, not suburbs.

The confidence ranges were just as broken. When a forecast claims 80% confidence — meaning the real outcome should fall within the predicted range 80% of the time — it needs to actually hit that 80%. Ours covered 46%. At the moments the model was most confident, coverage fell to 20%. The more certain the model was, the more wrong its confidence was.

We abandoned the forecaster entirely.

The insight that changed the approach: property booms are not events you catch before they start. They are multi-year processes you can catch early.

A suburb boom that runs for three years, detected six months in, still has the bulk of its growth ahead. The numbers from the backtest: entering a detected boom 6–12 months after it starts still captures 60–85% of total gains. The early-mover advantage of perfect prediction is much smaller than it looks — and the accuracy cost of trying to predict is enormous.

Boom size is also era-dependent: the median boom in the pre-2015 data was 1.3% growth above trend. Post-2020, that figure was 16.2%. In high-momentum periods, the cost of arriving late is lower than the cost of getting the wrong suburb entirely.

Detection asks a different and more tractable question: is this suburb booming right now, and how far through the boom is it? That question can be answered with current market data — price growth, days on market, vacancy rate, rental yield, and affordability relative to the city median. All publicly available. All suburb-specific.

The detection formula was validated against 78 suburbs — 28 that boomed, 50 controls that didn’t. Hard filters first: any suburb with annual growth below 5%, days on market above 45, vacancy above 2%, or a median price above $800K was excluded from scoring. Within the qualifying pool, five components combine into a score.

The separation gap matters as much as the accuracy number. Real booms score 20.2 points higher than false signals on average. The formula doesn’t just get the right answer — it gets it with margin. These aren’t borderline calls.

Zero false positives means no suburb the formula flagged as booming turned out to be a non-event. The accuracy drops are on the other side: the formula occasionally misses a boom that didn’t exhibit the expected market tightness signals. 85.7% leaves some booms undetected — it doesn’t generate ghost signals.

Detecting booms is one thing. The harder question: among all scored suburbs, do higher-scoring suburbs actually outperform lower-scoring ones? A walk-forward backtest across 12,360 postcode-months answers this.

Walk-forward backtest, 12,360 postcode-months. No lookahead. Excess return = suburb 12-month growth minus market median growth. Full methodology →

Perfectly monotonic across all four tiers. The top tier outperforms the market by 7.5 percentage points and beats it 71% of the time. The bottom tier underperforms by 6.4 points and beats it only 28% of the time. The spread from top to bottom is 13.9 percentage points.

This is the discrimination the forecaster couldn’t produce. Not which suburbs will grow in absolute terms — but which suburbs will outperform their peers within the same period.

Of all the signals tested, two survived every subsample and every version of the formula:

Both signals are public knowledge. The challenge is applying them across 8,417 Australian suburbs fortnightly, filtering the noise from thin markets where a handful of sales can distort the days-on-market figure beyond usability, and catching booms within weeks of starting rather than quarters.

That’s what the formula automates. Right now, 393 suburbs pass the hard filters — 35 under $400K, 149 under $600K, 204 under $800K — and receive fortnightly Strong Buy, Buy, Watch, or Pass labels based on their detection score.

The next time you read a property forecast — from a bank, a research house, a buyer’s agent — three questions are worth asking:

We publish the full backtest methodology — the 78-suburb detection validation, the walk-forward tier discrimination results, and the forecaster failure analysis — on our proof page. No gating, no email required. The maths is there to check.

The core finding: prediction at suburb level with publicly available data delivers coin-flip accuracy. Detection — asking whether a suburb is already booming and how early you are — delivers 85.7% accuracy with a 20.2-point separation gap and zero false positives across 78 real outcomes.

Mean reversion is real too. Past outperformers tend to underperform going forward. The suburbs with the strongest 5-year momentum showed no cross-suburb ranking advantage within any given period. The data consistently rewards buying into early booms in affordable suburbs — not chasing past growth.