French Property Prices 2025: Pipeline, Audit and Dashboard

A complete mission, run on public data so that every step can be shown: the 2025 property transfer file of the French tax administration (DVF), 3.8 million lines, turned into a controlled pipeline, an audited dataset, a DDI documentation and the dashboard above.

The deliverables are those of a client engagement: a layered pipeline with a rule catalogue, a quality report that explains what it set aside, an open table documented to international standards, and a dashboard anyone can question. Every figure can be recomputed: the code, the hash of the source file and the run log are kept.

• No sale identifier: one sale spreads over several lines (one per property unit or parcel) that must be grouped back.
• The declared value is repeated on every line of a sale: summing it inflates amounts.
• 301,443 lines are strictly identical to another one, and no unit identifier tells a duplicate from a second, identical outbuilding.
• Symbolic values, missing surfaces and outlying prices per m² distort any average.
• The data concern people: reuse must allow neither re-identification nor indexing by search engines.

• Bronze: the archive is kept unchanged, with its SHA-256 hash, size, row count and a check of the 43 expected columns. Address, parcel and lot columns are never parsed.
• Silver: typed values, INSEE commune codes (overseas departments included), duplicate lines set aside, quarantine with the reason for each rejected line, transfers rebuilt from date, nature, value, commune and disposition.
• Gold: a star schema ready for Power BI (sale facts, transfer facts, commune, department and date dimensions), kept on the processing machine.
• Publication: aggregates only, each price statistic resting on at least 11 sales and a commune median on at least 30; an open table by department; a quality report and a run log. The full run takes about 50 seconds.

Ten rules, each with an identifier, a level (file, line, transfer, sale) and an action: stop, quarantine, set aside, exclude from prices or flag for review. Their results are published with the dashboard: 301,443 duplicate lines set aside, 21,076 transfers without a usable value, 447 implausible or missing surfaces.

Outliers are detected with a robust z-score on the logarithm of the price per m², within each department and dwelling type: a flat at 9,000 €/m² is ordinary in Paris and exceptional in the Creuse. The median and the median absolute deviation resist the very outliers they look for. 14,727 sales (2.0 % of simple sales) are set aside this way; 4,614 more, three times above or below their commune median, are flagged for a person to review, not removed.

• Data minimisation: street address, cadastral parcel and lot numbers are never loaded into memory.
• Record-level tables stay on the processing machine: nothing finer than an aggregate leaves the pipeline.
• Statistical disclosure control: no price statistic under 11 sales, no commune median under 30 sales, values rounded to 10 euros.
• No third-party request: the charts are drawn by the site's own code, under a strict content security policy.

• A transfer is rebuilt from its values: two identical sales on the same day in the same commune are merged, and a transfer across several communes is split.
• House prices include the land.
• Off-plan sales often lack the description of the dwelling and are under-represented among simple sales.
• Bas-Rhin, Haut-Rhin, Moselle and Mayotte are not covered by DVF.
• The interactive view is built from aggregates: it never goes below the commune.