
The quality of loan-level data is a basic prerequisite for a well-functioning securitisation market. Investors, rating agencies and supervisors rely on it, for example in due diligence, valuation and ongoing monitoring. Errors or inconsistencies, however, are often detected late. Traditional checks rely on fixed rules, such as plausibility thresholds, or on sampling. A recent white paper by European DataWarehouse examines whether an unsupervised machine learning method, known as Isolation Forest, can usefully complement these checks.
How the method works
- Principle: Isolation Forest was developed in 2008 by Liu, Ting and Zhou. It splits the data step by step using random yes/no questions, such as “Income above X?” or “Loan amount below Y?”. A typical loan blends into the crowd and can only be separated from the others after many questions. An unusual loan, by contrast, stands on its own after just a few questions and is therefore flagged as anomalous.
- Unsupervised: The model does not need any pre-labelled cases of errors. This is an advantage, as confirmed errors or cases of fraud are rarely documented in securitisation data.
- Relative, not absolute: Each loan is given an anomaly score. The score indicates how unusual the combination of its characteristics is compared with the market as a whole. It does not indicate whether an individual value is wrong.
The study
- Data: The analysis covered UK RMBS transactions. It used numerical fields from the reporting templates, including borrower income, original and current loan balance, interest rate, property value, and original and current loan-to-value (LTV). Missing values and “No Data” entries were removed beforehand.
- Feature selection: Several model runs were used to test which variables contribute consistently to distinguishing loans.
- Results: Most loans form a dense core. The anomalies lie mainly at the upper end of the distribution, with high incomes, property values and loan amounts. Loans usually stand out not because of a single value but because of an unusual combination, such as a very small loan amount relative to a high income. For an individual transaction, the method can also show where its pool sits compared with the market and how much it overlaps with the anomaly clusters. The examples in the white paper are illustrative and are not based on real loan data.
What it means for the securitisation market
The approach is best suited as a prioritisation tool. In agreed-upon procedures, pool selection or due diligence, it shows which loans merit closer review. It uses only data that is already being reported. An anomaly, however, is no proof of an error or of weak credit quality. The results also depend on model assumptions, and the study is limited for now to one asset class in one country.
Outlook
Machine learning will not replace traditional data quality checks, but it can usefully complement them. The key questions are whether the method can be extended to other asset classes and markets, and how market participants and supervisors will deal with the results. If applied sensibly and developed further, the method can contribute to higher-quality data reviews and thus to the transparency and quality of the securitisation market.