Data Structuring and Analytics for a Credit Data Organisation
Non-profit association owned by member banks
The client is a non-profit association owned by its member banks worldwide. Its purpose is to help those banks understand and model their credit risk through data pooling and benchmarking. The value of the pooled data depended on resolving inconsistencies across contributors and on building analytics the members could use.
Scope of work
Probabilistic entity matching through a domain-specific model trained on large reference datasets from major market data providers. Disparate naming conventions and human subjectivity had made the same counterparty appear as several separate entities across source systems; the model removes duplicates and reproduces human judgement at scale.
Credit scoring from sparse data, using natural language processing with sentiment and topic modelling where credit risk and default data were thinly populated. The scores support pricing of illiquid credit and stress testing at single name and portfolio level, on a stochastic engine underneath.
Automated anomaly detection through a single name and portfolio analyser identifying both point and pattern anomalies across the pooled dataset.
Chained loss analysis, building unique identifiers from loan metrics and applying clustering to find common loss events contributing to losses at more than one member bank.
Early warning tool built on a semi-supervised model that learns from historical events and builds a database of topics and labels by country and sector:
- A proprietary distance approach applied across those topics
- Each obligor and its dependencies monitored continuously
- Recurrence of historically negative topics flagged against the name
Outcome and value
- A consistent counterparty view across data contributed by banks worldwide.
- Credit scoring made possible for names where member banks hold little default history.
- Loss events traced across institutions rather than examined bank by bank.
- Data and tools delivered together, so members could use the pooled set immediately.