Data Products & Projects

Clean, verified, provenance-tracked datasets and AI benchmarks for structured finance, company filings, and machine-readable financial infrastructure.

Home Projects About & Contact GitHub Hugging Face LinkedIn Academic Papers (SSRN)
Hugging Face Profile Licence Enquiry Article Back to Research
Daniel Cheah • Projects
Datasets Companies House Structured Finance LLM Benchmarks Hugging Face
Data Products

Licensable datasets and open benchmarks

These projects are designed around a simple thesis: models are increasingly commoditised, but clean, verified, provenance-rich financial data is scarce. The focus is not generic dashboards. It is analyst-grade data products that can be used by humans, LLMs, and downstream data pipelines.

UK Company Financials

Available for Licensing

An ML-ready dataset of 3.7 million UK company filings across approximately 3.5 million companies, parsed from Companies House iXBRL accounts and normalised across the FRS 102 and FRS 105 taxonomies.

  • Turnover, assets, liabilities, net assets, funding structure, and employees.
  • Per-figure provenance for auditability and model-training confidence.
  • Normalised UK GAAP fields, GDPR-clean design, and planned quarterly refreshes.
Dataset ↗ Pipeline ↗ Licence Enquiry

UK Securitisation SPV & Lender Map

Available for Licensing

A curated, structured-finance map of 551 UK securitisation SPVs and non-bank lenders, grouped by shelf, sponsor, and asset class, with charges-register intelligence converted into analyst-ready fields.

  • RMBS, specialist mortgage, auto ABS, card ABS, and equity release coverage.
  • Charges register mapping across trustees, fixed/floating security, and deal dates.
  • Designed for market mapping, originator analysis, and structured-credit research.
Sample & Details ↗ Licence Enquiry

SEC EDGAR Securitisation Connector & Skills

Open Source

A free, open-source Claude / Cowork plugin with a local SEC EDGAR data connector that turns US structured-finance filings into analyst-ready output — searching registered ABS and CMBS deals, pulling prospectuses and investor reports, and running loan-level analysis on Form ABS-EE data. Contributed back to Anthropic’s financial-services repository.

  • Local MCP connector (Python, standard library) exposing six tools; streams 100 MB+ ABS-EE loan tapes with flat memory.
  • Loan-level analytics for auto ABS and conduit CMBS: pool stratifications, balance-weighted coupon / FICO / DSCR / debt yield, property and geographic concentration, and the maturity wall.
  • The only no-subscription connector in the marketplace — free, public, rights-clean EDGAR data. Validated on real deals; Apache 2.0.
Pull Request ↗ Source ↗

UK Accounts LLM Benchmark

Open Benchmark

Can frontier LLMs read UK company accounts? A 1,000-question verified benchmark and five-model evaluation comparing proprietary and open-weight systems. The core finding: LLMs can read the accounts well; the defensible asset is the clean, verified data layer.

  • Benchmark coverage across Claude, GPT-5.5, Gemini, DeepSeek V4 Pro, and GLM 5.2.
  • Designed to show where model performance ends and data infrastructure begins.
  • Supports the commercial thesis behind the UK company financials dataset.
Benchmark & Code ↗ Read HTML Article →
Profile

Hugging Face

Dataset distribution profile for published samples, gated datasets, and machine-readable data products.

Open Hugging Face Profile ↗
Copyright & Proprietary Notice
© 2026 Daniel Cheah. All rights reserved. This material is provided for general information and discussion purposes only and does not constitute legal, financial, or investment advice. No reproduction, distribution, or commercial use without written permission.