The Dex Selection Backtest — Whitepaper (14 Years vs the Market)

$0.00

A forensic out-of-sample test of the HtAG Dex suburb ranking against every Australian house market across 14 anchor years, 2012 to 2025. Top-decile picks outgrew a price-band-matched benchmark by 8.85 percentage points over one year and beat it in 14 of 14 years. The paper also shows the edge is not a market-timing tilt (beta 1.01, alpha +10.65pp), reproduces its own benchmark from source data at r = 0.992, and states a ten-item limitations register plus the boundary of what the result supports in a credit or risk context. Written for analysts, risk reviewers and data-governance functions. Free.

Description

Description

Most property analytics providers assert that their ranking works. This paper puts HtAG’s Dex suburb-selection ranking on trial against the hardest benchmark available and publishes both what survived and what did not.

What was tested

At each historical anchor date the Dex selection signal was reconstructed from data available on or before that date, used to rank every Australian house market, and the top decile frozen as the pick set. Those picks were then scored against realised growth over the following 1, 3 and 5 years, against other markets in the same price band measured over the same forward window. A $450,000 pick is judged against other $450,000 markets, not against the national average.

What the paper reports

Results by horizon and price band. +8.85 pp excess at one year, +2.50 pp per year at three, +1.19 pp per year at five, with hit rates, downside percentiles and loss rates for each.
Year-by-year consistency. Positive in 14 of 14 anchor years, range +4.19 pp to +16.43 pp, with both the unweighted and the more conservative pick-weighted mean published.
The edge is not market beta. Beta to the universe 1.01, intercept +10.65 pp, and the excess uncorrelated with market conditions (r = 0.016, p = 0.96). The edge does not require a rising market.
The benchmark reproduces from source data. Rebuilt independently from HtAG’s locality price panel of 6,236 house markets: r = 0.992, mean absolute error 0.38 pp, mean bias zero, with the full year-by-year comparison table.
An honest treatment of sample size. Using the 14 anchor years rather than the 3,769 picks as the unit of observation, because picks overlap in time and cluster geographically: t = 12.50, p = 1.3 x 10^-8, bootstrap 95% CI +9.10 to +12.35 pp.
A claim withdrawn. Section 7 retires a previously published statement that risk falls monotonically with holding period, on the evidence of the same dataset that supports everything else.
A ten-item limitations register with materiality ratings, and a full data-lineage and point-in-time-status section.
Applicability to credit and risk decisions, split explicitly into what is supportable today and what is not, naming APS 220 and CPS 230.

Who it is for

Analysts, risk reviewers, model-governance and data functions evaluating whether HtAG’s outputs can be relied upon, and buyers agents or investors who want to understand what the selection evidence does and does not establish before acting on a ranking.

18 pages. Executive summary; two backtests, one word; what was tested; benchmark definition and validation; results; statistical treatment; a claim this evidence does not support; data lineage and point-in-time status; limitations register; applicability to credit and risk decisions; roadmap; conclusion; technical appendix; glossary.

Free download. Related reading: the Dex backtest results page, how HtAG validates its property data and models, and what backtesting is in property forecasting.