Methodology

How the model works,
and how we know it works.

Atlaso assigns growth ratings to 2,477 Australian suburbs using quantitative models trained on 3 million property transactions. This page describes the architecture, the validation protocol, the adversarial testing it has survived, and the limitations we know about. No number on this page is aspirational; every figure is reproduced by an automated test suite before each release.

The model

Three layers, deliberately simple:

• Direction — a blended logistic-regression and gradient-boosting classifier (25 engineered features for houses, 21 for units) estimates each suburb's probability of growing more than 8% over the next year. Features cover price dynamics, sales volume, vacancy, rents, asking prices, stock levels, affordability, and macro-credit conditions.
• Persistence — a rule-based structural quality score (yield level, vacancy level and trend, affordability, cohort momentum). No trained parameters; validated independently across 41 of 43 cities and 8 of 8 years.
• Value — a yield-based filter that gates the highest conviction tier: STRONG BUY requires a suburb to be cheap relative to its own history and its city, not merely trending.

Ratings: STRONG BUY, BUY, HOLD, AVOID. Suburbs with unreliable data (noisy quarterly medians, very thin sales, new estates) are barred from positive ratings regardless of score.

Validation protocol

Every published number is walk-forward: models are trained only on data available before each test year, then judged on what actually happened — repeated across ten years (2015–2024), 46 cities, one prediction per suburb per year. The validation applies the same data-quality gates the production system applies, so the numbers below describe the product as shipped, not a flattering variant.

Results (frozen 2026-06-11)

Measure	Houses	Units
STRONG BUY — beat inflation (>4% growth)	93.8%	97.1%
All picks — beat inflation (>4%)	77.0%	77.4%
STRONG BUY — grew >8%	90.2%	94.7%
STRONG BUY — average annual growth	+27.1%	+32.8%
AVOID — average annual growth	−10.7%	−19.5%
Direction ranking power (AUC, 1yr)	0.765	0.755

AUC with a one-year purge gap between training and test: 0.759 — the walk-forward is honest, not flattered by window overlap.

The long-term outlook score shown on suburb pages is separately validated against realised five-year returns (11,131 suburb-years): suburbs we label Strong returned +84.9% over the following five years on average, against +50.3% for Average and +22.0% for Weak — a strictly ordered ladder.

Adversarial testing

• Shuffled targets collapse the model to coin-flip (AUC 0.499) — no information leakage.
• Remove any single data source entirely and ranking power barely moves — no fragile dependency on one feed.
• Random half-samples of suburbs produce near-identical results (spread 0.003) — the signal is broad, not driven by outliers.
• 503 automated integrity assertions run before any release, covering reproducibility, leakage, gate logic, and per-city stability.
• Independently audited (June 2026): every published claim reproduced from raw data.

On the record, in advance

Data sources

Source	Coverage	Cadence
National sold-transaction records	3M+ transactions, 46 cities	Monthly
Vacancy, stock, rents and asking-price series	1,500+ postcodes	Weekly
ABS Census	2,600+ postcodes	Census cycle
RBA / macro credit series	National	Annual, publication-aware

All predictive features are lagged to what was actually knowable at prediction time — supply data by a full quarter, macro data by publication schedule.

The GREEN / RED regime framework

City-level regime badges summarise whether credit and market conditions currently favour growth. The suburb model is validated in both regimes: in GREEN years the market averaged +11.1% and our top-rated suburbs +18.0%; in RED years the market averaged +0.8% and top-rated suburbs +11.6%. The edge is the spread between what we rate highly and what we don't — in both halves of the cycle.

Limitations — read these

• We predict tiers, not magnitudes. "STRONG BUY" means historically high probability of outperformance, not "this suburb will grow 14.2%".
• All validation is historical. Past performance does not guarantee future results; the forward archive above is the live test.
• Average growth figures include boom years (2020–21). In flat markets the same ranking skill produces smaller absolute numbers.
• Suburbs with fewer than 3 sales per quarter cannot be rated at all; volatile thin markets are barred from positive ratings.
• Tier thresholds were calibrated on the validation window. Treat published hit rates as well-measured estimates, not promises.
• A suburb rating is not personal financial advice and cannot account for an individual property, contract, or circumstance.

← Back to Atlaso