Innovation Economy Index: Technical Whitepaper

1. Abstract

This paper presents a composite index that measures the concentration of innovative knowledge workers in Australian suburbs. The index combines census employment variables into a single score. Suburbs with clusters of knowledge workers in vibrant, fast-growing sectors score higher. Suburbs where these workers are largely absent score lower.

We trained a predictive model on historical property sales data to predict 2-year annualised growth rates. The model was evaluated out of sample and achieved an R-squared of 0.013. The top-scoring suburbs outperformed the national median by +1.9 percentage points over 2 years. The bottom-scoring suburbs underperformed by -2.9 percentage points.

The signal was tested across 187 quarterly periods, 26 individual sample dates, and 13 GCCSA regions. It held in 67% of quarters and was statistically significant at 19 of 26 sample dates. The spread was positive in 8 of 13 regions. But the signal breaks down after 2021 and inverts in 2022 to 2023. This paper documents both the strength and the failure modes.

2. Methodology

2.1 Feature Construction

The Innovation Economy Index is built from census employment variables that capture the concentration of innovative knowledge workers at the suburb level. We do not disclose the specific census field names. The combination represents proprietary research.

Each variable was standardised and combined into a single composite score using a weighted formula. The weights were derived from a model trained to predict 2-year annualised property growth relative to the national median.

2.2 Model Training

The model was trained on property sales data spanning 2008 to 2023. Each observation is a single property sale. The target variable is the annualised 2-year growth rate from the sale date. The features are census-derived employment variables for the suburb where the sale occurred.

The model was trained using a strict out-of-sample framework. Training and test sets were split by time period to prevent data leakage. Properties sold in overlapping windows were excluded from the test set to ensure clean separation.

2.3 Performance Metric

The primary metric is the difference in median annualised 2-year growth between each bin and the national median. Statistical significance is assessed using a two-sided t-test against the null hypothesis that the bin's mean growth equals the national mean.

2.4 Out-of-Sample R-squared

The model achieved an out-of-sample R-squared of 0.013. This means roughly 1.3% of variance in 2-year suburb-level growth can be explained by these census employment variables alone. Section 7 discusses why this level of explanatory power is still useful for investment decisions.

3. Bin Performance

The model sorts suburbs into three bins based on their Innovation Economy score. Each bin has a distinct growth profile. The table below shows the full results.

4. Temporal Analysis

A signal that works at one point in time could be a fluke. We tested the Innovation Economy Index across every quarter from 2008-Q1 to 2023-Q3. The chart below tracks the 2-year annualised growth rate for the above-threshold and below-threshold suburbs over time.

4.1 Date-by-Date Consistency

We tested the top bin's outperformance at 26 individual sample dates between 2012 and 2022. At each date, the model was evaluated independently. The result was statistically significant at 19 of 26 dates.

5. Regional Robustness

A signal that works only in one city is less useful than one that works nationally. We tested the Innovation Economy Index across all 13 GCCSA (Capital City Statistical Area) regions in Australia. The results are mixed.

5.1 Full Regional Table

All growth rates are annualised over 2 years. The spread column shows the difference between the above-threshold and below-threshold growth rates.

6. Suburb-Level Evidence

Suburb-level comparisons for selected cities are available on the summary page.

7. Defence of Method

7.1 Why a Low R-squared Still Matters

An out-of-sample R-squared of 0.013 means the model explains 1.3% of variance in 2-year suburb-level growth. This is low in absolute terms. It would be a poor result for a model trying to predict exact growth rates. But that is not how this index is intended to be used.

Property prices are shaped by hundreds of factors. Interest rate cycles, infrastructure investment, zoning changes, population growth, local employment, and macroeconomic conditions all play a role. No single thematic index will capture the majority of variance. A model that explained 50% of property growth from a handful of census and other government data variables would be implausible and almost certainly overfitted.

The relevant questions are different. Does the signal exist? Is it statistically significant? Does it persist across time? The evidence says yes, with important caveats about the post-2021 period.

7.2 Statistical Significance

The top bin's outperformance has a p-value of 2.5e-50. This is not borderline. The probability of observing a +1.9% difference across 112,564 sales by random chance is effectively zero. For context, a p-value below 0.05 is the standard threshold for statistical significance. The Innovation Economy Index exceeds this by 48 orders of magnitude.

7.3 Consistency Over Time (With Caveats)

The signal was significant at 19 of 26 sample dates. But unlike other Microburbs indices, the failures are not random. They cluster in two periods: 2018 (weak signal) and 2021-2022 (inverted signal). This is not noise. It suggests the relationship between innovation-economy employment and property growth is conditional on broader economic conditions.

When interest rates are stable or falling, suburbs with high concentrations of innovative knowledge workers tend to outperform. When rates rise sharply (as in 2022-2023), this relationship breaks down. Innovation-heavy suburbs may be more sensitive to rate rises because their residents carry larger mortgages relative to income.

7.4 Geographic Breadth

The spread is positive in 8 of 13 GCCSA regions. This is weaker than other Microburbs indices. The signal works best in regional areas (Rest of WA, Rest of Queensland, Rest of Victoria) and in Sydney. It fails in Melbourne and the ACT. This limits the index to specific geographic applications rather than national use.

7.5 Practical Use

This index works best as a secondary signal. It should not be the primary factor in suburb selection. When combined with other Microburbs indices that have stronger temporal consistency, the Innovation Economy Index adds value by identifying suburbs where knowledge-economy employment supports property demand. But it should be weighted lower than signals with broader geographic and temporal consistency.

8. Limitations

8.1 Post-2021 Signal Inversion

This is the most important limitation. The time series chart shows clearly that from 2022 onwards, the below-threshold suburbs outperform the above-threshold suburbs. The three most recent sample dates (2021-07, 2022-02, 2022-07) all show negative or inverted spreads. This is not a random failure. It is a structural shift.

The likely explanation is the 2022-2023 interest rate cycle. The Reserve Bank raised rates 13 times between May 2022 and November 2023. Suburbs with high concentrations of innovative knowledge workers tend to have higher house prices and larger mortgages. These suburbs were hit harder by rate rises. The very factor that drove outperformance in 2013-2017 drove underperformance in 2022-2023.

8.2 Narrow Feature Set

The index uses a narrow set of employment variables. This makes it less stable than indices built from six or more variables. A single structural shift in one employment sector can flip the entire signal. Broader indices are more resistant to this type of failure.

8.3 Government Data Is Point-in-Time

The index relies on government data sources including the Australian Census from 2016 and 2021. Census data is collected every five years. Employment patterns can shift between census dates. A suburb that scored highly in 2016 may have lost knowledge workers by 2021. The model cannot capture these shifts in real time.

8.4 Geographic Unevenness

The signal inverts in Melbourne (-0.40%), the ACT (-0.35%), Adelaide (-0.16%), and Darwin (-0.13%). Only 8 of 13 regions show a positive spread. Investors in Melbourne or Canberra should not rely on this index.

8.5 Low R-squared

The model explains 1.3% of out-of-sample variance. The remaining 98.7% is driven by factors outside this index. With a small number of input variables and a 2-year prediction horizon, this was expected. But it means the index should carry less weight than broader, more predictive signals.

8.6 2-Year Horizon

Unlike other Microburbs indices that predict 4-year growth, this index uses a 2-year horizon. Shorter horizons are noisier. A 2-year growth rate is more sensitive to short-term market swings, interest rate movements, and seasonal effects. This partly explains the lower consistency rate of 67% compared to 80% or higher for 4-year indices.

8.7 No Causal Claim

This paper documents a correlation, not a causal mechanism. We hypothesise that suburbs with high concentrations of innovative knowledge workers attract further investment, create local spending demand, and support property prices. But the data does not prove causation. The post-2021 inversion suggests the mechanism is more complex and condition-dependent than a simple causal story would allow.