Find housing markets that stand out from their surroundings — and see how reliable the comparison is.
Postcode Signals turns public house-sale records and geographic boundaries into an explainable area-level dataset for England and Wales. It asks: which postcode sectors have a median sold-home price more than 20% above nearby sectors?
The result can support geographic campaign experiments, market prioritisation and data analysis. It is a housing-price signal, not a measure of residents' income, net worth or purchasing intent.
Explore the example report · Download selected sectors · Product and architecture decisions · How it was built with agents
| September 2026 reference run | Result |
|---|---|
| Qualifying residential sales, 2023–2025 | 2,265,966 |
| Distinct sectors represented | 8,241 |
| Sectors with a usable two-step comparison | 8,064 (97.9%) |
| Sectors above the strict 20% local-premium threshold | 1,607 |
| Candidates remaining with at least 20 sales | 1,552 |
The original delivery retains 55 low-sale candidates with flags. The decision example uses a 20-sale floor and exposes sparse neighbours too. These are screening rules, not statistical confidence guarantees. Coverage and source evidence →
The maps distinguish locally higher prices, nationally expensive housing and insufficient sales evidence. The 20-sale floor belongs to the more conservative example shortlist; the original delivery retains sparse observations with flags. Figure notes and reproduction →
| Decision | How to use the output | Evidence still needed |
|---|---|---|
| Where to test advertising | Build candidate geographic test and control groups | Permitted targeting geography, data rights and measured incremental response |
| Which markets to investigate | Compare local premium, national percentile and sales counts | Serviceability, competition, demand and unit economics |
| What explains local variation | Join sector-level indicators and examine the graph | Property-mix controls, sensitivity checks and current geography |
Commercial activation requires checking upstream data rights and current platform rules. Aggregation alone does not establish permission for every use. See data terms and the experiment design.
Clone the repository and run from its root with Python 3.12:
python examples/area_shortlist.py --output reports/area-shortlist.md
python scripts/verify_snapshot.pyBoth commands use only the standard library and committed aggregate files. The first writes a readable shortlist report; the second checks the six published data files against their recorded sizes and SHA-256 hashes.
For the full CLI and tests, create a virtual environment and install the recorded dependencies:
python -m venv .venv
# Windows PowerShell: .\.venv\Scripts\Activate.ps1
# macOS/Linux: source .venv/bin/activate
python -m pip install -r requirements.lock.txt
python -m pytest -qTests exercise ingestion, geometry, graph traversal, exact selection boundaries and independent verification of the complete national output. CI runs on Linux and Windows; tests do not download national data.
flowchart LR
A["HMLR sales · 2023–2025"] --> B["Validate and aggregate sector prices"]
C["ONS Output Areas + sector lookup"] --> D["Dissolve approximate sector polygons"]
D --> E["Shared-edge adjacency graph"]
B --> F["Compare cumulative graph neighbourhoods"]
E --> F
F --> G["Select with price and sample-size filters"]
G --> H["CSV + decision report + quality flags"]
A small Python CLI keeps the decision inspectable: pandas for statistics, GeoPandas/Shapely for geometry and NetworkX for shortest-path neighbourhoods. There is no hosted service or database to operate. Source downloads are cached locally; compact derived outputs are versioned.
local premium = subject median / median(neighbouring sector medians) - 1
The neighbourhood includes distinct sectors one or two graph edges away and excludes the subject. Available sector medians receive equal weight; unpriced sectors remain in the graph for traversal.
Use a separate output directory to preserve the reference delivery:
python wealth.py prices --years 2023 2024 2025 --min-sales 20 --output-dir reports/run
python wealth.py geography --output-dir reports/run
python wealth.py analyse --layers 1 2 3 --output-dir reports/run
python wealth.py select --layer 2 --above 0.20 --min-sales 20 --output-dir reports/runUse --min-sales 1 on the last command to reproduce the original selection policy. The geography stage needs national downloads and local processing; it is not needed to read the CSVs. Publisher revisions can change a fresh rebuild. The original run_manifest.json is a fixed delivery record and is not regenerated by these commands. See methodology, exact sources and reproduction limits.
The initial delivery used seven behaviour-sized issues, separate implementation and validation commits, local tests and a national run. The build recorded GPT-6 Astra with xhigh reasoning, 14 commits and 34 passing tests at delivery. The current repository adds parallel review, regression fixes, CI and a reproducible decision example.
The development case study connects requirements to design tradeoffs, verification and the issue/commit trail. Exported usage counters record 1,833,042 tokens for the original implementation turn, including 1,722,880 cached input tokens. These counters are not unique text volume, cost or a productivity benchmark.
Prices describe sold housing, with no inflation, property-mix, floor-area or tenure adjustment. Approximate 2021 polygons do not cover every later postcode sector. Missing comparisons remain visible and cannot enter the shortlist. Commercial impact and individual wealth are unmeasured.
Original software and documentation: MIT. Data and derived figures: upstream attribution and terms. Contributing · Security reporting · Engineering backlog