Article · 3 min read

BCBS 239 data quality controls: green for two years, and an exposure missing from COREP

Stéphane Goethals
Stéphane Goethals
COO & co-founder of Vokse
LinkedIn
BCBS 239 data quality controls: green for two years, and an exposure missing from COREP

In a BCBS 239 framework, a data completeness check compares, every month, the number of rows in the warehouse exposures table with the number of records in the file delivered by the mainframe. It has been green for two years.

Upstream, a COBOL program silently drops loans whose currency code is missing from the rate table. The two counts match, since they are taken on a file that has already been filtered. The exposure is missing from COREP, and no check sees it. The same count, placed at the output of the original DB2 table, would have caught it at the first reporting date.

The check was not wrong. It was misplaced.

Why checks end up close to the report

In many banks, quality checks were placed where the data was visible: in the warehouse, at the input of the calculation engine, on the final report. It makes sense, since that is where the quality teams have access to the data. But a check placed downstream only sees what reaches it. Anything dropped, converted or rewritten before it escapes its measurement.

The BearingPoint survey of June 2026, covering 50 institutions in 9 countries, gives an idea of the gap: 24% report extensive documentation of their lineage. Without lineage, it is hard to know where a check should sit.

Linking each check to what it protects

Stéphane Goethals describes a continuous chain. For each data item, requirements. For each requirement, a check. Each check produces measurements, the measurements surface anomalies, and the anomalies go into an action plan with an owner and a date.

In practice, this means being able to answer three questions for every existing check:

  • which regulatory indicator it protects (a COREP cell, a liquidity ratio, an ICAAP indicator);
  • where in the chain it measures, and what happens before it;
  • who handles the anomaly when it turns red, and within what time.

When lineage is known end to end, the second question has a precise answer. You can then see the redundant checks, placed three times on the same data downstream, and the steps where no check measures anything.

What the supervisor looks at

The ECB’s RDARR guide expects controls along the whole in-scope data chain. In an inspection, the question is no longer “do you have checks”, but “show us that this check would have seen data lost upstream”.

A green check proves nothing until you know what it looks at.

See how Vokse links each check to the indicator it protects