YardPogo
YardPogo shows Nova Scotia listings and property records with plain-language model output alongside them. This page explains where every number on the site comes from and, honestly, how accurate each model actually is -- measured on real holdout data, not marketing copy.
| Data | Source |
|---|---|
| Active listings | CREA Data Distribution Facility (DDF), under licence |
| Sold prices | Nova Scotia public deed registry (property transfer tax return data, enabled by Bill 73) |
| Property facts (beds, baths, sqft, year built, assessed value) | Property Valuation Services Corporation (PVSC) open data |
| Model output (value estimate, time-to-sell, over-ask odds) | Our own models, trained on the above -- see below |
We do not display MLS-sourced sold prices or MLS listing history. Where a sold price appears on this site, it is the registry deed price, which is public record under Bill 73 -- not a number pulled from MLS.
A gradient-boosted model predicts a home's value from its property facts (bathrooms, square footage, age, style, garage), its assessed value, and its location (dissemination area, municipal unit, latitude/longitude). It is trained only on sales that closed before the year it is valuing -- a rolling, year-by-year holdout, so every accuracy number below is a genuine forward test, never hindsight.
Measured accuracy (rolling-origin backtest, 20 annual vintages, 2007-2026)
Each vintage is trained on nothing later than the year it prices and then scored against that year's actual sales. Per-year error ranged from about 11% (2009, 2019) up to 26.5% in 2021, when the pandemic-era market moved faster than any training window could track -- the model recovered to 11.9-13.4% by 2025-2026.
A competing-risks survival model estimates, for an active listing, the chance it sells within the next 7, 14, 30, 60, or 90 days, and a corrected median days-to-sell. It is trained on 3.97 million listing-week snapshots covering 259,484 listings, split by time: trained through December 2024, validated on 2025, tested on the first half of 2026.
Measured accuracy (holdout: Jan-Jun 2026)
Concordance of 0.651 (vs. 0.550 for a market-average baseline) means the model is meaningfully better than chance at telling which of two listings will sell sooner. Its predicted median time-to-sell, after a measured correction for long-sitting listings, comes out to 30.0 days against an observed 30.6 days across the same holdout. The uncorrected raw curve runs short in the long tail (homes still unsold after months are a harder-to-sell group than their day-one features suggested) -- we correct for that, we don't hide it.
A second model estimates the odds a listing sells above its asking price, and the expected premium (or discount) if it does, as a range rather than a single number. Trained on sales from 2012-2024, validated on early 2025, tested on February-July 2025.
Measured accuracy (holdout: Feb-Jul 2025)
A 0.732 AUC beats a coin flip (0.5) and beats the naive baseline of "assume the trailing neighbourhood median premium" (0.512 AUC, base rate 24.6% sold over ask). The premium estimate beats that same trailing-median baseline on error too (4.0% vs 4.51% mean absolute error). We calibrate the displayed percentage against real outcomes: before calibration the predicted-over-ask rate was off from reality by 2.9 percentage points on average (worst decile 6.0 points); after calibration that drops to 1.2 points (worst decile 7.5 points). We publish the after number because it's what actually ships.
A hazard model estimating the chance an active listing cuts its price in the next 7/14/30 days is in development and not yet live on listing pages. In backtesting on 2025 data it reached a 30-day AUC of 0.695 against a neighbourhood-cut-rate baseline of 0.444 -- promising, but until it clears our calibration bar for a live surface, any price-drop signal you see on the site is a plain cohort statistic ("X% of similar listings cut price within 30 days"), not a model prediction, and is labeled as such.
Where we show "comparable sales" behind a value estimate, they are selected using the same model that produces the value estimate, ranked by how similar their price-driving features are to the subject property -- not just distance and bedroom count. This is a differentiation project still in progress; until it ships, comps are chosen by the more conventional distance-plus-attributes method.
Hard limits, by design
Every model on this page gets re-measured on fresh holdout data as new sales come in, and we will publish a "how accurate were we" update on this page each quarter -- including when a number gets worse, not just when it improves.