How value is determined.

The most grounded read of value. For a listing, the site finds recent sales of comparable properties nearby — same property-type family (house / apartment / bungalow), within about one bedroom, within roughly ±15% of floor area, and in the same Eircode routing key (the first three characters, which is roughly a one-mile radius in Galway). If that yields too few, it widens the net rather than show nothing.

Each comp is inflation-adjusted to today's money using the CSO house-price index — a €200,000 Galway sale in 2014 is roughly €340,000 in 2026 money, and comparing without adjusting would make every older sale look cheap. The comps are then ranked by how close they are on distance, recency and size, so the most relevant ones surface first.

The headline comp figure is €/m² — price per square metre — taken as the median of the comp set (the middle value, not the average, so one odd penthouse can't swing it). The listing's own asking €/m² is compared against that median to produce the value score below.

Why thin data suppresses the score: only comps that actually carry a floor area contribute a €/m². If fewer than four do, the site shows no score at all rather than a confident- looking number resting on two sales. This is deliberate — it's the difference between “we don't know” and a false bargain.

The second, independent read. An AVM (automated valuation model) is a model trained on roughly 14,000 Galway sales to learn what each feature is worth: floor area, beds and baths, BER, property type, year- built bucket, and location. Hand it an active listing's features and it returns an estimated price.

It's a gradient-boosted tree model (LightGBM), chosen over a neural network because 14,000 rows is a small dataset where a tuned tree model gets within striking distance for a fraction of the complexity. It trains in seconds. Prices are learned in log space and inflation-adjusted, the same as the comps.

Crucially the AVM returns not just a point estimate but a band — a low (p10), middle (p50) and high (p90) estimate — so you see the model's uncertainty, not a single falsely-precise figure. The honest limits: it knows nothing about condition (a gutted comp and a renovated one look identical to it), and location signal is coarse, which is what the next step fixes.

The AVM values a property from its features but has no sense of street or estate identity. Two houses with identical features on different roads get the same prediction, even when one estate consistently trades 8% above the model and the next trades below it. That gap is real, local, and visible in sales that already happened.

So before trusting the model's number, the site checks how wrong the model has recently been nearby: it looks at the model's error on recent sold properties close by, weights them by distance and recency, and nudges the prediction by that local bias. If the model has been running 6% low around here, the estimate shifts up about 6%. When few neighbours exist the correction shrinks toward zero — it only trusts a pocket once enough sales agree.

Measured on the backtest, this moved the model's median error from 10.14% to 9.89% — a modest, broad gain, biggest where the plain model was weakest (rural and county stock, where estate premium is strongest and the model's location signal is coarsest).

A price band is only useful if it means what it says. An 80% band should contain the real sold price 8 times in 10 — and the point is that we measure whether it does. The raw AVM band was too tight: its nominal 80% band only actually covered about 70% of sold prices, so “looks under-priced” flags fired on noise.

The fix is conformal calibration: the band is widened by a data-driven margin, learned from how often recent sold prices fell outside it, until its measured coverage matches the 80% target. Because the model works in log space, the widening is proportional — it stretches a €150k cottage's band and a €1.2M house's band by the same relative amount, not a flat euro figure.

The honest cost: a truthful 80% band on this data is wide — roughly 48% of the price, versus the old tight (but dishonest) 39%. The old band bought its tightness by under-covering. One known gap remains: the cheap / rural under-€250k tail is still slightly under-covered, because it's genuinely the hardest segment to value.

The value score you see on a listing is the comps read: how far the asking €/m² sits below (or above) the median €/m² of comparable sales. A positive score means asking is below the comp median — potentially under-priced; negative means above it. It's deliberately simple and interpretable: “below half of similar properties' price per square metre” is plain English.

The AVM is the separate model read. The two are computed independently — comps from raw nearby sales, the model from learned feature prices — so when they agree, that's genuine corroboration. When they disagree sharply, the listing carries a “comps and the model disagree” warning telling you to dig in, rather than picking one and hiding the other. Showing both, and their agreement, is the whole design.

Confidence is tied to the real evidence: the count of comps that actually carry a €/m² (8 or more is high, 5–7 medium, fewer is low), and a wide spread in the comp set caps confidence down. The score is never reported without that sample size attached.

A listing is flagged as looking under-priced when its asking price sits below the model's calibrated low estimate (p10) — i.e. cheaper than the model's 80% band, now that the band is honestly wide. But a low price has two very different explanations, and the site refuses to conflate them:

• A genuine deal — a motivated seller, an off-season listing, a thin market for that exact property.
• A discounted lemon — a poor BER, 110 days on the market, a third price drop. Correctly priced for its problems, not under-priced.

So a flag is a prompt, never a verdict. Every listing shows the residual signals next to the score — BER, days on market, the count and size of price drops — so a 12% discount “explained” by an F-rated BER and 110 days unsold reads as exactly what it is. The human (you) decides.