A residual land value model that asks one question per development: can a developer
pay these owners more than they will accept, and still make its margin?
Working, not validated. Every completed sale in the test set is a success —
there are no failures in the sample — so the model can be checked for accuracy but not for
false positives. Treat the ranking as a shortlist for investigation.
Read the probability, not the score. The residual-land-value model returns a
ratio. A score of 2.6 is not a 260% chance of anything. Every table below now carries
a calibrated probability of a completed collective sale within five years,
and the number that governs all of them is the base rate: .
The best development in Singapore on these features sits at —
high relative to the base rate, still unlikely in absolute terms.
Methodology
The model works backwards from what a completed project can sell for, subtracts everything
that must be paid before owners see a cent, and compares what is left against what owners will
accept. Nothing is scored on "feel".
Revenue runs on floor area, not unit count. A developer maximising units on an
underutilised site is slicing the same cake thinner, not baking a bigger one. Measured across 804
within-project matched pairs, smaller units command no PSF premium at all (ratio
1.000) — the apparent gradient across size bands is a composition artefact. Chuan Park is the
clean illustration: GFA rose 5% while unit count rose 105%, because average size fell from 1,390
to 909 sqft. Unit count still matters for absorption risk against the ABSD clock, which
this model does not yet price.
achievable selling PSF from launch pricing within 3km
↓
land budget = maxLandPsf(sell) × redevelopment GFA
− Land Betterment Charge on the GFA uplift only
− Lease Upgrading Premium leasehold sites, Bala's Table to a fresh 99 years
− ABSD 5% non-remissible; the 35% is remitted upfront
↓
MAX PAYABLE TO OWNERS ÷ existing units
vs
OWNER REQUIREMENT = market value × 1.45
↓
ECONOMICS = payable ÷ required × CONSENT (age tier, owner count) = SCORE
The cost stack is measured, not assumed
Rather than guessing a construction cost, the land-to-breakeven relationship is regressed
from 66 real launches that carry both a land price and a published breakeven:
breakeven PSF = 339 + 1.5228 × land PSF ppr R² = 0.948, n = 66
realised margin over breakeven: median +25.4% (n = 37)
Planning envelope
Site area and zoned plot ratio come from URA Master Plan 2025 parcel geometry, joined by
point-in-polygon to 1,771 developments. The dwelling-unit cap follows URA circular DC22-10:
max dwelling units = (MP allowable GPR × Site Area) ÷ 85 sqm
÷ 100 sqm in nine named localities
no cap at all in the Central Area
Validated against a known outcome: Chuan Park computes 910 units against the
916 in Kingsford's actual scheme — 0.7% out.
Consent thresholds
Age tiers follow the Land Titles (Strata) (Amendment) Bill introduced 4 August 2026 —
90% under 10 years, 80% for 10–39, 70% for 40–59, 65% for 60+.
The Bill simultaneously tightens the requisition bar to 35%, halves the signature window to six
months, and extends the post-failure wait to three years, so its net effect is not assumed to
be positive.
From ratio to probability
The economics model says whether a deal is possible. It cannot say how likely one is,
because likelihood depends on how rare collective sales are in the first place — and they are
very rare. That base rate is the anchor, and everything else moves a development around it.
posterior odds = base rate odds x c x ( PRODUCT of likelihood ratios ) ^ lambda
base rate completed sales per year / developments old enough to be candidates
LR_i P(feature | sold) / P(feature | universe), measured per factor
lambda correlation damping — DU headroom and floor-area uplift correlate +0.70,
so multiplying them as independent evidence double-counts
c intercept, solved so the mean probability across the universe returns
to the base rate. Factors REDISTRIBUTE likelihood; they cannot invent
deals the register does not show happening.
Why not a logistic regression
A logit needs failures to separate from successes, and this sample has none — the register
records completed sales only. Likelihood ratios can be measured against the whole universe as
the negative class instead, which is what this does. It is a weaker instrument and the numbers
below are stated with that in mind.
The weights, and where each one comes from
There are no chosen weights. Each factor's weight is its measured likelihood ratio:
the share of completed sales in a bin over the share of the universe in the same bin. A ratio
above 1 raises the odds, below 1 lowers them, and 1.00 means the factor carries no information.
Bins are smoothed toward the universe distribution, so a bin with no completed sales in it
returns a mild penalty rather than a verdict of impossible.
How wide the uncertainty really is
Every ratio above rests on 14 reconstructed sales. Resampling those 14 with
replacement 400 times, rebuilding the ratios and the calibration from scratch on each draw,
gives the bands shown beside each probability. At the top of the list the band spans roughly a
factor of three. That width is the honest precision of this exercise, and it is the reason
these numbers belong in an internal shortlist rather than in front of an owner.
Factors that earn their place
What we tested and rejected
Widely repeated criteria that do not survive contact with the data.
Rejected
MRT proximity. 86% of sold en blocs sit within 800m of a station — and so do
87% of all 30-year-old condos. Lift 0.98×. Tightening to 500m or 300m does not
rescue it: at every radius the base rate falls inside the confidence interval. Median distance
is 550m for sold sites against 510m for the universe. Location is already priced through the
selling-price input; adding it again would double-count.
Rejected
Freehold preference. 12 of 14 sold sites are freehold, but the two largest
deals — Chuan Park ($890m) and Loyang Valley ($880m) — are both 99-year leasehold with ~55
years left, precisely what the advice says to avoid. The skew is a size artefact. The only
published study finds tenure insignificant (p = 0.18 and 0.95).
Rejected
Acquisition-cost ceiling. Above $1bn: 12% of candidates, 14% of sold — lift
1.19×. Loyang Valley and Chuan Park both exceeded $1bn all-in and both sold. Acquisition cost
also correlates +0.90 with unit count, which is already in the model.
Demoted
Plot ratio gap as a gate. Real, but not a requirement. Median measured uplift
is 1.38×, five of fourteen are below 1.2×, and Chuan Park sold at 1.05× — already
built to 96% of its zoning. Density and land value are substitutes (corr −0.46), so the model
treats them as alternative paths rather than a filter.
Adopted as a floor
Dwelling-unit headroom. URA caps units at GFA ÷ 85 sqm (÷ 100 sqm in nine
localities, no cap in the Central Area). Comparing that cap against units standing today:
no completed sale had headroom below 1.5×, against 18% of candidates, and
29% of sales sat above 4× against 10% of candidates — lift 2.94×. Real, unlike MRT.
Applied as a floor rather than a weight: it correlates +0.70 with GFA uplift, which
already drives the economics, and with n=14 the independent 30% cannot be separated from noise.
⚠ The false signal: high unit headroom with shrinking floor
area. Honolulu Tower shows 5.27× DU headroom on 0.65× GFA uplift — a developer could slice
thirty huge flats into many normal ones, but has 35% less area to sell, and smaller units
earn no PSF premium (measured 1.000 across 804 within-project matched pairs). 41 candidates
carry this pattern and are flagged.
Adopted
Unit count. The strongest survivor. Median completed deal 2021–26 is
22 units; 21 of 24 are under 150. Significant at p = 0.00 in both published
logit specifications.
Validation against completed sales
For each completed collective sale, the model's maximum payable to owners against the price
actually paid. Inputs are the site's real parcel geometry and the caveat record.
0.92
Median predicted ÷ actual
13 / 14
Within ±40%
+83%
Median realised premium to owners
The economics of one deal, line by line
Pick a completed sale to see the whole arithmetic: what the developer paid, what came off the
top before owners saw anything, what each owner received against what their home was worth, and
what a new unit on the site has to fetch for the numbers to work.
What every one of these deals needed a new unit to fetch
The five sales with enough pre-deal caveats to price them share one finding, and it is the
model's binding defect. Each site had to command roughly three times its own resale
PSF as a new launch. The scorer's fallback rule for a site with no launch comparable
assumes 1.35×.
This is a test of the fallback, not of the live model: all
scored candidates price off launch comparables within 3km, so the
fallback is currently dead code. It is shown because the multiple is the number that matters —
live candidates are priced at a median 1.74× (freehold) and 2.27× (leasehold) of their own
resale PSF, against the 2.92× these deals actually required. The model is reading the selling
price low across the board, not only at Loyang Valley.
These sites cannot be rescored as of their sale dates: per the survivorship
correction, only two of them still geocode, because the buildings are gone. Not out-of-sample
either — the Land Betterment Charge rate was calibrated on Loyang Valley, which appears here.
The probability the model would have assigned
The fairer test of a shortlist is where a development that really sold sat in the
distribution. 50% is a coin.
Top 20 by score
Ranked from developments aged 30+ with a readable parcel, after
excluding commercial and mixed strata and developments under 10 units. Ranking is driven by
economics — payable ÷ required — not by either uplift ratio.
⚠ marks sites already built to or beyond their zoning.
A redevelopment there yields less floor area than stands today, so it can only work on
land value alone. This is the Horizon Towers pattern — the most famous failed collective sale in
Singapore was over-built at 3.13 against a 2.8 plot ratio. Treat flagged rows with particular
suspicion.
exPR existing plot ratio · GPR zoned ·
GFA uplift = GPR ÷ exPR, the ratio that drives revenue ·
Unit uplift new units ÷ existing units, a consequence of the URA dwelling-unit cap
rather than a source of value (at Chuan Park these are 1.05× and 2.05×) · Econ payable ÷ required, ≥1 means a deal can
clear · Acq total acquisition cost including ABSD, BSD, LBC and lease premium ·
p(5y) probability of a completed collective sale within five years, with its
bootstrap band.
Ordering is now by probability rather than by score, which moves rows: a site can
carry excellent economics and still rank below a weaker one that sits in better-evidenced bins
on owner count and floor-area uplift. Click a row to see which factor contributed what.
Where the leasehold candidates went
The top of the list is entirely freehold. That is a real property of the output and it is
worth being precise about the cause, because the obvious explanation is the wrong one.
Not this
A freehold preference in the model. There is none. The tenure multiplier was
removed after EdgeProp's logit found tenure insignificant (p = 0.18 and 0.95), and the
likelihood ratio measured here is — statistically indistinguishable
from no effect. Median economics are freehold against
leasehold. On the economics itself, tenure does nothing.
Cause 1
Owner count, which travels with tenure. Median freehold candidate:
units. Median leasehold candidate: .
Singapore's old freehold stock is boutique walk-up blocks; its old leasehold stock is large
HUDC-era estates. The consent curve is calibrated to a register whose median deal is 22 units,
so it penalises size hard — median consent factor freehold against
leasehold. The freehold skew is an owner-count skew wearing a
tenure label.
Cause 2
The Lease Upgrading Premium. Topping a site back up to a fresh 99 years costs
a median of the entire land budget, and it comes off the top
before owners see a cent. Freehold sites pay nothing. This is a real cost, not a modelling
choice — Loyang Valley's was $246m — but it means a leasehold site must clear a bar its
freehold neighbour never sees.
Cause 3
Seven "leasehold" developments are 999-year leases and are freehold in every
way that matters — Roxy Square, Le Loyang, Kensington Park, Grande Vista, Ridgewood, Allsworth
Park, Azalea Park. They were inflating the leasehold count while behaving as freehold. They are
now classified as freehold, which is why the true 99-year leasehold pool is
developments and not 48.
The counterfactual: leasehold with the lease premium removed
Setting the Lease Upgrading Premium to zero — not a proposal, a diagnostic — shows how much of
the leasehold gap it accounts for on its own.
Several perennial candidates cross 1.00 the moment the lease premium comes out,
which locates the constraint precisely. It does not make them viable: the premium is genuinely
payable. What it shows is that leasehold sites are not failing on developer appetite or on
location — they are failing on a statutory cost plus an owner count the consent curve punishes.
The best true-leasehold candidates
The strongest leasehold development ranks overall,
and only appears in the top 50. Against the two largest completed
deals in the register both being 99-year leasehold — Chuan Park and Loyang Valley — that ordering
should be treated as a caution about the consent curve's steepness, not as a finding that
leasehold estates do not sell.
Live attempts
Developments currently launched for collective sale. These are the prospective test set — a
failure is worth as much as a sale, because it is the only way to measure false positives.
Worked example — Lakeside Towers, asking $350m
144 units on 153,237 sqft, GPR 2.1, 99-year lease from ~1979, third attempt. Existing plot
ratio ~1.81 gives just 1.16× uplift, and the lease top-up premium takes $59–91m
off the top before owners see anything.
The ask implies $968 psf ppr, close to the $1,077 psf ppr its neighbour
Lakeside Apartments achieved in May 2022 — which is presumably how it was set. But that neighbour
was less built-out with a longer lease. On residual value the ask needs about
$3,112 psf to work, against roughly $2,500 for the nearest launch. Every completed
deal cleared at 70–143% of model max payable; this ask is around 135–210% of it.
Known limitations
No failure data. All 14 test cases sold. False positives are unmeasured, so
the probabilities are calibrated against the universe as a stand-in negative class rather than
against attempts that failed. Live attempts resolving is the only fix.
The base rate's denominator is asserted, not measured. Roughly 1,000 private
non-landed developments aged 30+, from the Bill's own coverage figures. Every absolute probability
on this page scales inversely with that number; the ordering does not depend on it at all.
The register is stale at the recent end. Coverage by year is shown below.
2014 holding zero deals is genuine — the market was shut between the 2013 TDSR framework and the
2016 restart. 2024 onward is not: reporting describes around five completed sales in 2025 against
the one carried here, and Thomson View ($810m, 255 lots, 99-year lease from 1975,
agreed Nov 2024 and completed 2025) was absent until it was added by hand. It is the largest
leasehold comparable available and its absence was actively skewing the tenure evidence.
Thomson View is in the register but not in the outcome reconstructions. That
needs its total strata area from a REALIS caveat pull; without it the existing plot ratio cannot
be measured, and estimating it would put a guess into the one table that is meant to be measured.
The selling-price input reads low across all five measurable deals, not just
Loyang Valley — they needed a median 2.92× their own resale PSF, against 1.74–2.27× that live
candidates are being priced at. This depresses every economics figure on the page, and it
depresses large leasehold sites most because their statutory costs are fixed and come off first.
Land Betterment Charge is one calibrated rate ($4,300/sqm of uplift, tuned to
Loyang Valley) standing in for SLA's real 118-sector table. It is extrapolated well beyond its
fitting range on high-uplift sites.
Selling-price comps miss en bloc redevelopments. The launch file covers
Government Land Sales sites only, so a redevelopment like LakeGarden Residences — the single most
relevant comparable for Lakeside Towers — is absent.
Four of nine URA "100 sqm" localities cannot be mapped to any published
boundary; those are approximated and flagged low-confidence, affecting unit yield by ~15%.
Central Area has no unit cap in reality; the model falls back to the 85 sqm
quantum as a neutral assumption for 223 developments.
Premium floor rests on five projects. Measured, not invented — but not yet a
stable constant.