How it works, and where it can be wrong

The core idea: house prices as a random walk

House prices are modeled as a lognormal random walk (geometric Brownian motion) -- the same assumption used for stock prices:

HouseValue_P(t) = Price0 × exp(t·μ + z_P·σ·√t)

μ and σ are the mean and standard deviation of the log-return of the index (ln(index_t / index_t-1)) -- not of the raw year-on-year percentage change. Volatility scales with √t, not t, so a multi-year band widens more slowly than a naive fixed-annual-rate projection would suggest.

Where μ and σ come from: a four-tier fallback

Every growth estimate used anywhere on this site comes from the most granular level with enough real data, falling back automatically:

1. District -- a district's own recent price-per-square-foot trend from real NAPIC transactions, blended with its region/state estimate via Buhlmann credibility: Z = n / (n + K), K=8. Only 188 of the possible district × property-type combinations have enough transaction volume (≥15 transactions/year, 2021 onward) to get their own blended estimate.
2. Region -- NAPIC's sub-state index, published for 12 of 14 states (not Kelantan or Perlis), and not every property type within a state.
3. State -- MHPI's official annual index, all 14 states, 2009-2025.
4. State's "All House" -- the final fallback if a specific property type isn't separately published for that state.

K=8 is a chosen round number ("8 comparable annual observations for 50% credibility"), not fitted to this data's own variance -- deliberately conservative, since NAPIC's usable transaction history only spans 2021-2025 (at most 4 year-on-year observations per district), so no district ever gets more than ~33% weight on its own trend. Sigma is never credibility-blended -- there isn't enough district-year data to trust a variance estimate at that level, so it always falls back to the region/state figure.

A project's own trend is not used, and what was tested

The projection starts from the area's growth (the four tiers above), fades it toward the long-run national rate, and blends it half and half with what the same kind of homes in the area actually resold at over the last three years. A single project's own trend used to replace the area's when the project had enough sales. It no longer does. The blend with recent resales applies to the starting growth only: the long-run rate the growth fades toward stays the national one, so a recent boom or slump in an area shapes the first few years but does not carry into the 30-year figure.

Why: we tested it on about 20,000 same-size unit kinds across 3,700 projects, forecasting from the year before and scoring against what really happened. The central path with the project's own trend missed by 15.4 points over two years; without it, 12.9. A project that rose was no more likely to rise again (the link was slightly negative), because one project has too few sales a year for its own trend to be more than noise.

Also tested, not adopted: starting from the last two years of the official index instead of the full-history rate cut the error on the index alone, but did not beat the current blend with recent resales (two-year error 5.3 points against 4.9 on area averages), and under-forecast Johor Bahru. ARIMA and other time-series methods did worse than the simple rules.

Limits: the project-level test only had sales from 2021, so three starting years. Places that speed up suddenly (Johor Bahru and Penang's mainland in 2024 and 2025) were forecast too low by most methods, and no method built on past prices can see a shift like that coming. The ranges are wide for that reason. Full results are in docs/PROJECTION_MODEL_STUDY.md.

Owing more than the home is worth: what we looked at

NAPIC records have no unit IDs, so we cannot follow one home or one mortgage. We rebuilt cases instead: every recorded sale from 2021 to 2024 is a buyer, in a kind of home (project, tenure, type and size) with at least three sales in a recent window. Each is assumed to have borrowed 90% of the price at 4.5% over 30 years and paid on time, and to pay 3% to sell. Other loan sizes and rates are tested.

Two ways to date "today", because the open file is uneven. The NAPIC open file holds about half of the residential transactions NAPIC reports for 2023 and 37% for 2024, and from the fourth quarter of 2024 only about a third as many sales a quarter as before, unevenly (Selangor, Johor and high-rise lose most). So we ran the study twice: with "today" as the latest four quarters (53,512 buyer cases, 1,195 projects), and as the last four full-coverage quarters, 2023 Q4 to 2024 Q3 (118,053 cases, 4,009 projects). Where they differ, both are shown. We also compare with a strict measure (what that kind of home sold for against what the buyer paid, which counts overpaying) and a market measure (the project's price change since the buyer's year), and with how many cases chance alone would produce.

What holds in both. Across the market, 1.7% to 3.5% of buyers would owe more than the home sells for at a 90% loan, about what chance alone gives (2.4% to 3.1%). The size of the loan matters a great deal: at a 100% loan about 13% in both windows. Paying well over the typical price for the kind of home in the project is the clearest personal risk: of those who paid 15% or more over, 36% to 48% are in a home worth less than their loan, against 3% to 5% of those who paid about the typical price. Most owners whose project fell are still above their loan (84% to 92%). The cases are concentrated: 20 projects hold 53% to 74% of the buyers who would owe more than their home is worth, out of 760 to 1,860 projects, and which projects they are changes with the window.

What does not. Claims about a kind of home or a place swing with the window (serviced apartments 4% to 17%, Kuala Lumpur 3% to 10%, launch-year buyers 2% to 14.5%), mostly because a few large projects dominate each, so we make none. For the same unit kinds seen in both windows, serviced apartments were about flat and KL flat, not collapsing, while terraced houses rose about 4% to 6%.

Limits: assumed loans and recorded prices, not real mortgages or real sales. Only 2021 onwards, with no downturn like 1997 or 2008 in it. NAPIC may include developers' first sales, which we have not confirmed. What goes with a loss is not the same as what causes it. Method and tables are in docs/NEGATIVE_EQUITY_STUDY.md.

Renting a home out: what can and cannot be known

There is no open data on what a particular home rents for (NAPIC's rent tables are in its paid Property Market Report), and the yield tables on property blogs could not be traced to a source, so we use none of them. The rent is your own figure. What we can test is whether tenants could pay it: a tenant should spend no more than 30% of income on rent, and household income in each state and district (Household Income Survey 2024) follows a lognormal curve that we fit from the median and the Gini coefficient. As a check, the average income the fit implies is within 10% of the survey's own for 92% of districts. The share of households earning enough is an upper bound on the tenant pool, because renters earn less than the average household, and it is a ceiling: a tenant can usually find a cheaper comparable home, which our data cannot see.

For the earnings, the price follows the same simulation as the Buy vs. rent page (the two give identical house values for the same seed), rent and costs grow with simulated inflation, and the comparison is a fixed yearly return you pick (low-risk savings, a fixed deposit, Malaysian shares or your own). Published average rents (Juwai IQI) against the median sale prices in our data give gross yields of roughly 4% to 6.5% depending on place and kind of home: a consistency check, not a finding, because the mix of homes differs. Full method and limits are in docs/RENTAL_STUDY.md.

How easily a home might sell: what the sales data can and cannot say

NAPIC's open file records completed sales, not listings: no time on market, no homes that failed to sell, no asking prices. So we cannot say how likely or how fast one particular home sells. What we can count is how many comparable homes (same project, tenure, type and size band) changed hands, and how well that pace predicted another sale in the following year. Tested out of sample on 77,479 kinds of home (pace over 2022 Q1 to 2023 Q3, outcome 2023 Q4 to 2024 Q3, and again on a second pair of periods): kinds selling under 1 a year saw another comparable sale in the next 12 months 16% of the time, 1 to 2 a year 44%, 2 to 4 a year 69%, 4 to 8 a year 85%, 8 or more 92% to 94%. Sales come in clumps, so even busy kinds had a quiet year 12% of the time.

Most kinds of home sell rarely: 63% sold under once a year (detached 91%, high-rise 42%). Where the number of homes is known (45 researched developments, mostly leasehold flats and condominiums in Kuala Lumpur), recorded sales were a median of 1.7 per 100 homes a year, and the file holds only about half of the market, so true turnover is probably about twice that. We use 2022 and 2023 because the open file is thin since late 2024 and a quiet recent spell in it does not mean homes stopped selling. Details and limits are in docs/LIQUIDITY_STUDY.md.

Work city or home town: how the two places are compared

Three ways to end up in your home town are run through the same simulation: buy where you work and sell when you move; rent there, save what buying would have cost, and buy at home when you move; or rent there and buy at home now. All three share one monthly budget (what buying where you work would cost each month) and the same starting cash, and are compared at the year you move, with the home town home counted as owned at its market price after every cost. The work-city price, savings return and inflation follow the Buy vs. rent page draw for draw, so the first two paths are exactly that page's buying and renting. The home town price is a second path with its own growth (the same central path as everywhere), moving partly with the first: the yearly correlation of the two states' official house price indexes over 2010 to 2025, shrunk toward the average pair by the share of the spread between pairs that is sampling noise (each pair rests on only 16 yearly changes). Details and limits are in docs/TWO_PLACES_STUDY.md.

The correlated Monte Carlo (buy vs. rent page)

House price growth, low-risk savings returns and CPI inflation are simulated together, using a 3×3 correlation matrix (Cholesky decomposition to generate correlated shocks), because in reality they move together. That correlation matrix is estimated from only 10 overlapping annual observations (2016-2025) -- treat it as indicative, not precise.

The buy-side net worth (house value minus remaining loan) has an exact closed form and is computed live for any input, never precomputed. The probability that buying beats renting-and-investing genuinely needs simulation, because it depends on the joint distribution of two different, correlated random processes over time -- there's no exact formula for that. Every run on this site uses your exact inputs; nothing shown is a precomputed default scenario silently standing in for your actual numbers.

A probability that rounds to 100% or 0% in a finite-sample simulation is still not a certainty -- it means zero of however many thousand simulated paths crossed over, at that specific spread. This site caps displayed headline probabilities at >99% / <1% for that reason, and always shows the growth-rate-vs-mortgage-rate spread that's actually driving the result next to it.

Save and buy in cash (no loan)

For people who want no interest-bearing loan at all. Each simulated future draws a house-price path (the same model as above), a path for the return on your investment and a path for your salary. Every year the pot grows by that year's return and a share of that year's salary is added; the first year it covers the price plus stamp duty and legal fees, you can buy. Pay figures come from the Department of Statistics Malaysia's Salaries & Wages Survey (citizens, 2010-2024). Your starting salary is the typical pay for your job role. Pay then rises by the market trend for your industry (the rise in its typical salary between the 2013-15 and 2022-24 averages, less the part that is only the workforce getting older, with the yearly spread measured leaving out the COVID and Census-rebasing years) plus the raise that comes with age, read from how typical pay differs between age groups. Raises for age are never negative, because the lower pay of older age groups mostly reflects how they started out rather than pay cuts. You stop saving at 60, the minimum retirement age. This is a typical-worker trend, not a forecast for one person's career. The low-risk return is the same one used on the other pages; any other investment is your own assumption, and whether it meets your shariah requirements is for you and your adviser to confirm. Rent, food and other costs while you save are not modelled -- they are covered by the share of salary you leave unsaved.

Transaction costs (Malaysia, current as of Sept 2026)

MOT stamp duty follows the tiered scale (1% / 2% / 3% / 4% across price bands). Loan agreement stamp duty is a flat 0.5% of the loan, as the banks' own fee sheets state. Legal fees follow the Solicitors' Remuneration Order scale plus 8% SST. First-time Malaysian citizen buyers get a full exemption on both stamp duties if the price is ≤ RM500,000 and the SPA is executed between 2026-01-01 and 2027-12-31 -- a time-limited policy, not a permanent one. These rates are policy-driven and move with the national Budget -- re-verify before launch and periodically after. Real estate agent commission is out of scope, since it's paid by the seller on a future resale that this model doesn't simulate.

What this model doesn't know

• Rental yield, income, DSR ceiling and wait-period savings rate are all your own inputs or illustrative assumptions -- not derived from real Malaysian rental or income data. They're labeled inline wherever used.
• The correlation between house prices, low-risk savings returns and CPI is a 10-year estimate -- short enough that it should be treated as directional, not precise.
• District-level credibility blending is deliberately capped for the same reason: a few noisy years of transactions shouldn't override a 16-year official index.
• This tool doesn't model a future resale, agent commissions, renovation costs, or the review/reputation of a specific development -- those are explicitly out of scope for this version.