An Automated Valuation Model (AVM) produces an indication of a property's value from data and statistics, in seconds, without a physical inspection. Abroad, banks, funds and property portals rely on them daily; Cyprus has had almost none built on its own transaction record. This article summarises a working AVM for the Cyprus residential market, developed as an MSc Data Analytics thesis and now serving two purposes: an internal evidence tool for valuers, and a public estimate. Throughout, the design follows the same logic a RICS valuer applies in the Sales Comparison Approach — value a property by what similar properties actually sold for — but does it at scale.

The data foundation. The model is built on the Department of Lands and Surveys (DLS) Comparative Sales record: fifteen official workbooks covering all five government-controlled districts from 2011 to 2026, roughly 219,500 completed sales. The data was sourced and cleaned independently, end to end, so that every cleaning and encoding decision is transparent and defensible for both professional practice and academic scrutiny. Two categories are kept deliberately apart — completed transfers, used to train the model, and contracts of sale, held back as recent market evidence rather than training material.

The Land Registry's remarks field on each sale. A core contribution is treating the Land Registry's free-text remark field — attached to every sale — as structured evidence rather than noise. A purpose-built Greek-language parser folds accents (accent-folding), reads area abbreviations (enclosed space, covered and uncovered verandas), and even resolves handwritten sums such as "143+4=147". A second, rule-based step then classifies each transaction. This matters because DLS records the same asset in different ways: a flat is booked cleanly, but a standalone house is often booked under its land, with the building's floor area left blank — that size had to be recovered from a separate enrichment layer. Overlook this quirk and the true house market is understated by roughly half.

Data enrichment. That layer draws on a public DLS interface (API) returning each unit's characteristics — year built, physical condition, quality class, view. These fields are sparse in the raw export, populated on under half of flats; after enrichment, coverage rises to around ninety per cent, which is what makes quality and condition usable as value drivers at all.

Three markets, three models. Flats, houses and building plots behave like three different markets, so each is modelled separately — per district, across five districts. (The occupied north is excluded.) Per-district modelling matters because Limassol's coastal-luxury segment behaves nothing like Nicosia's domestic market; forcing one national model to average across them harms both. Each model reads a fixed set of fourteen property attributes (features). For flats, internal floor area alone explains about a third of the result; for houses, quality class and built area lead; for plots, where there is no building, location and the planning zone's building density dominate.

The modelling choices, in plain terms. Four decisions keep the model grounded. First, it predicts the logarithm of price, so a handful of very expensive coastal sales cannot dominate the arithmetic. Second, it returns a range of values rather than a single false-precision figure: a lower and an upper price (the 5th–95th percentile band) within which the true price is expected to fall about nine times out of ten — and that range widens exactly where the market is genuinely harder to price. Third, economic logic is enforced on the model: a property in poor condition, for example, can never be estimated higher than an otherwise identical property in good condition — even if, in some small local pocket of sales, the data happen to suggest otherwise. Fourth, because sales take twelve to twenty-four months to register at the Land Registry, historical prices are adjusted to today's market and the estimate is nudged forward to the current quarter. Every estimate is also cross-checked against the nearest genuine comparable sales.

How well it works. Performance was measured with leakage-safe 5-fold cross-validation — the data is split into five parts, and each time the model is scored only on sales it never saw during training — against the international IAAO assessment standards. In plain language: MAPE is the typical percentage error, the coefficient of dispersion (COD) measures consistency, and the price-related differential (PRD) measures whether the model is equally accurate on cheap and expensive properties. Flats are strongest: in four of the five districts the model typically lands within about nine to twelve per cent of the eventual sale price and passes all three standards; Limassol, the largest and most varied flat market, is the single exception, sitting just outside. Houses perform well overall, with residual error concentrated in standalone coastal villas, whose value turns on land and bespoke features the Land Registry never records. Building plots are the hardest of the three: the land market is thinner and more varied, so estimates are wider and lean more heavily on the uncertainty range.

Transparency and limits of use. The tool is deliberately framed as an indicative estimate, not a RICS Red Book valuation. Where a segment is known to be less reliable, the interface leads with the range and a plain warning; unusually large plots and standalone houses are routed to a qualified valuer rather than given a machine figure. Missing data is declared, not invented — bedroom counts, absent on two-thirds of records, were left out rather than guessed. The result is decision support with its workings on show, not a black box: a valuer can inspect the comparable evidence behind every number and judge it.

For Cyprus, the wider point is that a defensible, data-driven valuation model can be built on the country's own public record — provided the messy detail of how that record is kept is respected rather than glossed over.

Try the model live at avm.axiavaluers.com. 

Sources

• Department of Lands and Surveys (Republic of Cyprus) — dls.moi.gov.cy
• RICS Valuation – Global Standards (Red Book) — rics.org
• IAAO, Standard on Ratio Studies — iaao.org

Trifonas Mamas
Property Valuer (MRICS–ΕΤΕΚ)
Registered Estate Agent
MSc Data Analytics Candidate