European DataWarehouse (EDW) recently published a whitepaper exploring how Isolation Forest can be used to identify unusual patterns in securitisation data and place them in a wider market context. Read on to discover the key findings and what they could mean for the future of data-quality analysis.

Within large loan-level datasets, individual values can appear perfectly plausible while the combination of characteristics surrounding a particular observation is statistically unusual.

That creates a challenge for data-quality analysis.

Traditional rules are designed to identify known inconsistencies. But what about unusual observations that have not been defined in advance?

A new EDW whitepaper explores exactly that question, testing whether an unsupervised machine-learning technique called Isolation Forest can complement existing data-quality checks by identifying observations that appear unusual relative to the wider securitisation market.

EDW already performs data-quality checks on information uploaded to its database to identify possible errors and inconsistencies. The research explores whether Isolation Forest could complement that existing set of verifications.

The distinction is important.

Instead of searching only for observations that violate a predefined rule, Isolation Forest seeks to identify unusual observations by progressively isolating records through random partitions. Records requiring fewer partitions to isolate are considered more anomalous, while observations embedded in dense groups of similar records are considered normal.

For the pilot, the EDW team applied the methodology to eligible UK RMBS transactions and tested numerical fields from its securitisation repository database. These included Primary Income, Original and Current Principal Balance, Current Interest Rate, Current and Original Loan-to-Value and Current Valuation Amount. Missing values and No Data entries were removed from the analysis.

The objective was not merely to identify extreme values.

The research found that anomalous observations were not driven by a single variable such as interest rate. Instead, the algorithm considered multidimensional relationships between characteristics including income, collateral value and loan size. That matters because each individual characteristic may be plausible on its own while the overall combination is uncommon within the wider dataset.

EDW plotted anomalous observations against the wider UK RMBS population and then highlighted observations belonging to a particular transaction. This provides an immediate indication of where a transaction sits relative to both typical market behaviour and observations identified as statistically unusual.

The visualisations showed the anomalous population occupying clearly defined regions of the broader market distribution. For the selected transaction analysed in the paper, observations were concentrated in the upper range of the core market population and partially overlapped with anomalous clusters across several variable combinations. The result is therefore more than an automated outlier list. It provides a benchmarking perspective, showing how individual loan characteristics and transactions compare with a much broader market population.

There is an important caveat.

An observation being statistically unusual does not mean that it is incorrect. Nor, as the whitepaper makes clear, should an anomalous classification on its own be interpreted as evidence of fraud, misreporting or credit weakness. Instead, the results provide a starting point for more targeted review.

That distinction is critical when applying machine learning to complex financial datasets. The algorithm helps identify where to look. Subject-matter expertise is still needed to understand what the observation means.

The initial results demonstrate the applicability of Isolation Forest to EDW’s consistent and standardised loan-level data, identifying statistically unusual loans and comparing them with the wider market.

The next step is scale.

According to the whitepaper, the framework can be automated within EDW’s analytical environment and is ready to be extended across other asset classes, enabling repeatable anomaly screening across a wider transaction universe.

For EDW, this research represents another way of examining the considerable volume of standardised loan-level information available within its repository. And it points towards an interesting evolution in data quality: from checking what we know to be wrong, to identifying what looks unusual enough to deserve a closer look.