Inspiration
HDB resale buyers make major financial decisions using mostly backward-looking information. We wanted to see whether Singapore's open housing data could provide a useful signal about where prices may be heading next.
What it does
HomeSignal forecasts median HDB resale prices 1–6 months ahead by town and flat type.
A buyer can compare historical prices, expected price movement, and eventually an affordability benchmark.
Unlike a typical dashboard that answers "What happened?", HomeSignal aims to answer "What may happen next?"
Our working forecasting model already beats a no-change persistence baseline at every 1–6 month horizon, reducing MAE by about 14–19%, with a best improvement of 19.17%.
Code and model pipeline:
https://github.com/ADante08/flatfair
How we built it
We use official HDB resale transaction data and the HDB Resale Price Index from data.gov.sg.
Transactions are aggregated into monthly segments by:
month × town × flat type
We engineer historical price and market features and train separate XGBoost models for each 1–6 month forecast horizon.
Our backtesting is time-aware: each historical prediction only uses information that would have been available at the forecast date.
For the full product, we plan to use Databricks for:
Lakeflow Jobs → Delta + Unity Catalog → Bronze/Silver/Gold → MLflow → forecast tables → AI/BI Dashboard or Databricks App
Challenges we ran into
Our first models struggled to consistently beat a simple persistence baseline. Adding more complexity did not solve the problem.
The breakthrough came when we realised that local price history alone was missing the broader movement of the HDB market. We incorporated the HDB Resale Price Index (RPI) as a market-level feature, which helped the model capture overall price trends alongside town- and flat-type-specific behaviour.
We also had to be careful with time-series leakage, making sure each historical forecast only used information that would actually have been available at that point in time.
Accomplishments that we're proud of
The forecasting core is already working.
It:
- forecasts 1–6 months ahead;
- beats persistence at every horizon;
- improves MAE by about 14–19%; and
- produces forecasts by town and flat type.
What we learned
We learned that forecasting is not just about choosing a model. Time-aware validation, leakage prevention, meaningful baselines, and reproducible experimentation are just as important.
We also saw how Databricks can help move the project from a local ML experiment into a maintainable data product.
What's next for HomeSignal
If selected for the build sprint, we plan to:
- automate the pipeline with Lakeflow;
- manage data with Delta Lake and Unity Catalog;
- use MLflow to track and improve model experiments;
- add household-income data for affordability analysis; and
- build a polished AI/BI Dashboard or Databricks App.
Genie Code can also help accelerate data exploration, debugging and dashboard development.