Combine models and correct residuals
Improved the AVM with a weighted ensemble and residual correction, increasing R² and within-10% accuracy.
CASE STUDY / MICROBURBS · AUTOMATED VALUATION MODEL
More reliable Australian property valuations
An Australian property valuation model combining weighted ensembles, residual correction, and confidence estimation to make predictions more reliable.
01 / THE PROBLEM
What needed to work.
An automated property valuation model needs to account for differences between properties, locations, and market conditions. Alongside improving prediction performance, this work addressed how to estimate the reliability of a valuation.
02 / MY APPROACH
How I approached it.
Represent local market behaviour
Engineered spatial, temporal, and market-driven features, including price momentum, regional price dispersion, historical volatility, and locality-level aggregates.
Estimate prediction confidence
Designed a confidence estimation model using quantile bootstrap regression to evaluate the reliability of the AVM’s predictions.
03 / SYSTEM OVERVIEW
Connecting the pieces.
- 01Property & market features
- 02Weighted ensemble
- 03Residual correction
- 04Valuation & confidence estimate
The work combined predictive accuracy with reliability. Feature engineering represented variation in local markets, while the confidence model added a separate view of how much trust to place in a predicted value.
AUSTRALIAN PROPERTY / AVM02
+12.47%
Model performance improvement
Weighted ensemble+Residual correction
04 / OUTCOMES
Results in context.
AVM performance
Overall performance improvement reported in the résumé.
Above Domain’s AVM
Reported comparative performance; evaluation protocol is not provided in the résumé.
Public overview of my contribution at Microburbs. Metrics are reproduced from the résumé; no additional test-set sizes, baselines, or evaluation results are implied.