Wiki · glossary

Every term, defined.

One plain-English definition for every term the site uses. For the full story on how the value figures are built, see how value is determined.

A–Z

23 terms

Ask vs sold: How far a property's final sold price landed from its asking price, as a percentage. Positive means it sold over asking (a bidding war); negative means it was negotiated down. Reported per area and per month.
AVM: Automated Valuation Model — a model trained on past sales that estimates a property's price from its features (area, beds, BER, type, location). The site's AVM returns a low–middle–high band, not just one figure, so you see its uncertainty.
BER: Building Energy Rating, A1 (best) to G (worst) — Ireland's mandatory energy label. A poor BER is one of the strongest reasons a listing is cheap for a reason rather than a genuine deal. A pending or exempt rating is treated as unknown, not as low.
Canonical listing: When the same property appears on several sources, the site keeps one row as the canonical (shown) listing and marks the rest as duplicates of it. Daft wins by default because it carries the most data; the others appear as “also listed on”.
Comps: Comparable sales — recent sales of similar properties nearby (same type, similar beds and floor area, same Eircode area), inflation-adjusted to today's money. Their median price per square metre is the most grounded read of what a listing should cost.
Conformal calibration (CQR): A method that widens the model's price band by a data-driven margin so its real coverage matches the target. It's how an 80% band is made to actually contain the sold price 8 times in 10, rather than the ~70% the raw model managed.
Coverage: The percentage of real sold prices that fall inside the model's predicted band. An honest 80% band should have ~80% coverage; the production model measures about 79.5%. Anything well below the target means the band is too tight to trust.
Days on market: How long a listing has been live without selling. A high figure (say 90+ days) with no movement suggests the market thinks it's mis-priced — context that turns an apparent discount into a warning.
Dedup: Deduplication — collapsing the same property's multiple source listings into one canonical row. It matches on Eircode plus a bed/house-number check, or a strict fuzzy address match, tuned to never wrongly merge two different homes.
Eircode + routing key: An Eircode is Ireland's seven-character postcode (e.g. H91 ABC1). Its first three characters, the routing key (H91 = Galway city), define a roughly one-mile area used to find nearby comps. Eircodes are often missing in older PPR data.
Harness / backtest: The test rig that scores any valuation method against real past sold prices, strictly time-ordered so the model never sees the future. A change ships only if it beats the current model's numbers here — it's the gate, not a vanity metric.
MdAPE: Median Absolute Percentage Error — the headline accuracy number. It's the middle of all the model's percentage errors, so half its estimates are at least this close to the real price. The production model's MdAPE is about 9.8%.
Methods agree: Whether the two independent reads of value — comps and the model — point the same way. When they agree it's real corroboration; when they disagree sharply the listing is flagged “treat with caution” rather than trusting one in silence.
p10 / p50 / p90: The model's low, middle, and high price estimates — the 10th, 50th, and 90th percentiles. The p10-to-p90 span is its 80% band; p50 (the median) is its robust point estimate. An asking price below p10 is what triggers the “looks under-priced” flag.
PPR: Property Price Register — Ireland's official record of every residential sale since 2010, filed from stamp-duty returns. It's the ground truth for sold prices, but lags by weeks and carries only price and date, not features like floor area.
Price per m²: A property's price divided by its floor area in square metres — the standard way to compare homes of different sizes. The value score compares a listing's asking €/m² against the median €/m² of its comps.
Repeat-sales: A valuation approach (Case-Shiller style) that tracks the same property selling twice to measure pure price change. The site rejected it: Galway's ~38,000 PPR sales over 15 years have too few repeats, and it produces a market index, not a per-property value.
Residual correction: An adjustment that checks how wrong the model has recently been on nearby sales and nudges its estimate by that local bias. It recovers street and estate premium the feature-based model can't see, helping most in rural areas.
RPPI / time adjustment: The CSO's Residential Property Price Index, used to restate old sale prices into today's money before comparing them. Without it a €200k sale from 2014 (~€340k today) would make every older comp look artificially cheap.
Sale agreed: A property where an offer has been accepted but the sale hasn't legally closed. It doesn't yet appear in any sold dataset — that only happens after conveyancing completes and it files on the register, typically one to three months later.
Value score: The headline comps figure on a listing: how far its asking €/m² sits below (positive) or above (negative) the median €/m² of comparable sales. It's suppressed entirely when fewer than four comps carry a usable price, so thin evidence never reads as a confident bargain.
Walk-forward: The leak-free way the model is tested: for each month, it's trained only on sales from strictly before that month, then scored on that month's actual sales. This prevents it from “predicting” the past using the future, which would flatter the numbers.
Within-10%: The share of the model's estimates that land within 10% of the real sold price — an intuitive accuracy hit-rate alongside MdAPE. The production model puts just over half (about 51%) of its estimates within 10%.