FlatFinder

Inspiration

Choosing an HDB resale flat means making a major financial decision with information scattered across transaction records, income statistics and market reports. A buyer comparing two towns can see what flats sold for, but it is harder to judge what those prices mean for their budget or how uncertain future prices are. We created the FlatFinder concept to make that comparison clearer.

What it does

FlatFinder will let buyers explore resale prices by town, flat type, storey range and year. They can compare similar past transactions, see a price-to-income measure, and view a six-month price forecast range. Each forecast will show its sample size and flag areas with limited data.

The income measure compares a median resale price with annual median household employment income. It is a national benchmark, not a personal affordability calculation. So it does not account for grants, CPF use or loan terms.

How we built it

We have designed the buyer journey, selected the datasets and mapped the Databricks pipeline; the dashboard and model are planned for the two-week sprint.

We will use HDB resale transactions, SingStat household employment income and resident population by planning area, all from data.gov.sg. Lakeflow Jobs will ingest the data into Delta tables governed by Unity Catalog. We will check changing schemas, missing values, duplicates and unusual prices before preparing town-level summaries. MLflow will track rolling forecast tests, while Databricks SQL and AI/BI will power the buyer dashboard. Genie and Databricks Apps are possible extensions if time allows.

Challenges we ran into

The main challenge in designing FlatFinder was deciding what makes a comparison fair. A town’s median price can hide differences in flat size, storey and remaining lease. Some town and flat-type groups may also have too few monthly sales for a dependable forecast.

The supporting datasets have limits too: household income is a national benchmark, while the population data is a 2020 snapshot organised by planning area rather than HDB town.

Accomplishments that we're proud of

We’re proud that FlatFinder grew from a question we might face ourselves: where could we realistically afford a home? We thought about what we would want to know as buyers: which towns fit our budget, how prices compare, and how much we can trust a forecast. Those questions helped us turn a big housing issue into a tool with a clear purpose and goals we can test during the sprint.

What we learned

A useful forecast needs more than a predicted price. Buyers should see how many transactions support it and where the model has limited evidence. We also learned to describe the price-to-income measure carefully: it helps compare market prices, but it does not include grants, CPF use or financing.

What's next for FlatFinder

In the two-week sprint, we plan to build the cleaned data pipeline, interactive town comparison and six-month forecast. We will test whether its mean absolute error is at least 10% lower than a seasonal-naive baseline, then ask five pilot users to try shortlisting a town. These are goals we have yet to measure.