Data & methodology

How we build the dataset

1. Discovery. We read the Housing Ombudsman’s public XML sitemaps to find every published decision and landlord page, and use each entry’s last-modified date to crawl incrementally.
2. Polite crawling. We fetch the server-rendered HTML directly, identifying our crawler, honouring a crawl delay and capping concurrency. We respect robots rules and cache raw HTML so re-runs do not re-fetch.
3. Parsing. From each decision page we read the title (landlord name + case reference), the one-line complaint summary, and the cleaned body text of the determination.
4. Structured extraction. A language model reads each determination and records outcomes, complaint categories, order types, themes, compensation and a plain-English summary against a fixed schema, with a confidence score. We mirror the Ombudsman’s own determination language and never assert beyond the stated finding.
5. Entity resolution. Landlords are matched to a canonical registry built from the landlord sitemap, collapsing name variants and (where available) cross-referencing the Regulator of Social Housing for homes-owned figures used in per-10,000-homes rates.
6. Indexing. Decisions are stored in a database and indexed for fast search, filtering and league tables.

What the metrics mean

• Maladministration rate — the share of a landlord’s decisions with an adverse finding (maladministration, partial or severe).
• Findings per 10,000 homes — adverse findings normalised by the number of homes a landlord owns, for fairer comparison between large and small providers.
• Total compensation — the sum of compensation the Ombudsman ordered landlords to pay residents across decisions in scope.

Limitations

Automated extraction can misclassify or miss detail, and the published archive only covers cases the Ombudsman has determined and published. Treat the structured data as a research aid, not legal advice, and confirm anything material against the original determination.