What impacts HDB resale prices most?

A model-and-verification study of 234,303 registered transactions from 2017-01 through 2026-06. The July 2026 partial month is excluded.

1. The ranking

I used three tests and combined their ranks: grouped permutation (scramble a whole driver family), leave-one-family-out retraining, and how much each family predicts on its own. This guards against a common error: calling one variable unimportant merely because a correlated variable can substitute for it.

Permutation MAE increase: how much the full model worsens after scrambling that family. Leave-out MAE increase: how much a freshly trained model worsens without that family. Standalone R²: how much log-price variation that family explains alone. These quantities are not additive.

Inside the bundles: the month is the strongest single field

When each raw field is scrambled separately, transaction month is the strongest individual field. Within location, the exact block-and-street address carries much more signal than the broad town label. Within size, measured floor area carries far more signal than the room-type label.

2. What “comparable flat” differences are worth

To isolate practical premiums, I compared transactions within narrow cells: same month, town, flat type, model, storey band and lease band for floor area; analogous cells for lease; and same address/year/type/model/area band for storey. These are observational estimates, not guaranteed valuation rules.

3. Timing and location are the two big contextual forces

The official HDB Resale Price Index rose 51.5% from 2017-Q1 to 2026-Q2. That is a market-wide shift, not a feature of any individual flat. Location premiums below come from a separate recent-period hedonic model that holds month, flat type, model, area, lease and storey constant.

4. Verification: the result survives harder tests

• Random holdout: tests prediction on unseen transactions while every month is represented in training.
• Out-of-time holdout: trains through Dec 2024 and predicts Jan 2025–Jun 2026 after deflating prices with the official RPI. This tests whether structural relationships persist.
• Alternative model: a sparse one-hot ridge model checks that the result is not an artefact of nonlinear gradient boosting.
• Subgroup stability: rankings are recomputed across time periods and major flat types.

5. How to use this

• For buyers: benchmark within the same month, town or address, flat type, area and remaining-lease band. A national median is a poor anchor.
• For sellers: do not overprice a high-floor or special-model premium; those are smaller than size/location/timing effects in the verified comparisons.
• For valuation: use the model as a screening benchmark, then inspect renovations, condition, orientation, MRT/school proximity and transaction-specific circumstances—the official dataset does not contain them.

Data and methodological limits

HDB notes that transactions depend on many factors and excludes some non-market cases such as resales between relatives and part-share transfers. This analysis explains registered prices; it does not prove causal effects. Address is a powerful location proxy but does not separately identify MRT distance, schools, view, orientation, noise, renovation quality, ethnicity quota constraints, urgency, or buyer/seller bargaining. Remaining lease can also affect financing eligibility, so its association includes both consumption value and financing-market mechanisms.

Sources: HDB resale transactions; HDB Resale Price Index; HDB Property Information. Code, raw downloads, model object, metrics and charts are included in the accompanying analysis bundle.