Data Quality Framework for Banks in Six Steps

Keep Data Quality High Across Many Systems

Only 2 of 31 global banks fully met Basel's risk data rules in 2023. A data quality framework fixes errors where one system feeds the next.

Alexander Kiel

Business Analyst · Banking & Wealth Management

10.07.2025 · 6 MIN READ · PROCESS OPTIMISATION

ON THIS PAGE

1. Scope the Data Quality Framework

2. Agree One Definition

3. Name an Owner and a Golden Source

4. Measure Against Clear Rules

5. Reconcile Between Systems

6. Fix It at the Source

Sample case: Citigroup

In short

FAQ

Poor data is expensive. Gartner estimates that it costs organisations at least $12.9 million a year on average. When Thomas Redman and his colleagues asked 75 executives to check 100 recent records from their own departments, on average 47% of newly created records had at least one critical error. Only 3% of the scores reached an acceptable level. A data quality framework turns that cost into owned, measurable tasks.

Banks are no exception. In its 2023 review of 31 global systemically important banks, the Basel Committee found only two banks fully compliant with all of its risk data principles. It linked many of the weaknesses to a fragmented IT landscape and legacy systems.

The problem rarely sits in one system. It sits between them. Does your bank know which system holds the right address, risk profile or domicile of a client?

STEP 1

Scope the Data Quality Framework

A bank holds millions of data fields across core banking, CRM, portfolio management, payments and reporting systems. Not all of them need the same care. Start with the few that drive regulatory reports, risk figures and client outcomes, such as client identifiers, domicile, risk profile, positions and prices.

Supervisors expect this focus. The ECB guide on risk data aggregation asks that the critical data elements underlying the key risk indicators should also be explicitly identified. The Basel Committee also urges banks to ensure sound data quality as the basis for digitalisation projects, not as an afterthought.

✓

List the reports and decisions that matter most

✓

Trace each one back to the data fields it uses

✓

Mark these fields as critical data elements

Which twenty data fields would cause the most damage in your bank if they were wrong?

Keep in mind: A list of hundreds of critical data elements is no longer a priority list. Start small and extend it later.

STEP 2

Agree One Definition

Systems often use the same word for different things. "Client" can mean a person in CRM, a contract in core banking and a household in portfolio management. "Domicile" can mean tax residence in one system and postal address in another. Until the definitions match, the data cannot match.

The Basel Committee's principles state that a bank should have a "dictionary" of the concepts used, such that data is defined consistently across an organisation. Ten years on, its 2023 review still found that some banks lack a common taxonomy and complete data lineage.

✓

Write one business definition for each critical data element

✓

Record allowed values, formats and the system fields that map to it

✓

Publish the glossary where every team can find it

Would three departments in your bank give the same definition of an active client?

Keep in mind: A definition agreed only by IT is not enough. The business teams who use the data must sign it off.

STEP 3

Name an Owner and a Golden Source

When a value differs between systems, someone must decide which one is right. For each critical data element, name one business owner and one golden source: the system where the value is created and maintained. Every other system copies from there. In the projects I have worked on, systems disagreed most on client static data such as domicile or risk profile.

The ECB expects data owners responsible for key risk indicators and critical data elements across the whole process. Thomas Redman describes the roles as data creators and data customers. Customers should tell creators what they need. My article on reference data in a bank migration shows how one golden source per domain works in practice.

✓

Name one accountable owner per critical data element

✓

Define one golden source system for each element

✓

Stop manual edits of copied values in other systems

When two systems disagree on a client's domicile, who in your bank decides?

Keep in mind: Ownership sits with the business, not with IT. IT runs the systems, but the business knows what the data should say.

STEP 4

Measure Against Clear Rules

What is not measured is not managed. Turn each definition into rules that can be checked automatically. The UK Government Data Quality Hub uses six dimensions: accuracy, completeness, uniqueness, consistency, timeliness and validity. Consistency matters most when the same data lives in several systems.

Yet Gartner reports that 59% of organisations do not measure data quality. The ECB asks banks for data quality indicators with tolerance levels and robust correction processes. A simple scorecard per data element makes the results visible to management.

✓

Write measurable rules for each dimension and data element

✓

Set a tolerance level and an owner for each rule

✓

Report results monthly, with trends rather than snapshots

Could you show your management the error rate of your ten most important data fields today?

Keep in mind: A quality score without a threshold invites debate instead of action. Agree what "good enough" means before the first report.

STEP 5

Reconcile Between Systems

Each interface between systems is a point where data can be lost, delayed or changed. Compare the golden source with every system that copies from it: record counts, key values and totals. Investigate each break and close it, not only the large ones.

The Basel Committee's principles state that risk data should be reconciled with bank's sources, including accounting data where appropriate. The ECB adds that banks should maintain complete and up-to-date data lineages, so each figure can be traced back to where it came from.

✓

Reconcile critical data between the golden source and each copy

✓

Document the data lineage from capture to report

✓

Track open breaks with an owner and a due date

How many manual spreadsheets sit between your source systems and your regulatory reports?

Keep in mind: Spreadsheets used to fix reports are systems too. Include them in the lineage and in the reconciliation.

STEP 6

Fix It at the Source

Cleaning data downstream treats the symptom. The same error returns with the next delivery. Redman argues that most efforts go into cleaning up errors rather than finding and fixing the root cause. He calls for creating data correctly the first time.

The ECB also expects a root cause analysis and clear processes for remediating and escalating data quality issues. Many root causes sit at the point of capture: free text fields, missing checks or manual re-entry. My article on mapping the process before you change it shows how to trace an error back to the step where it starts.

✓

Analyse the root cause of each recurring data break

✓

Add validation rules where data is entered

✓

Remove manual re-entry between systems

Which errors does your team correct every month without ever asking where they come from?

Keep in mind: A fix at the source often needs a change in another department's process. Agree it with that owner, not only in your own team.

What's So Dangerous About Bad Data? | Georgina Sturge (Data Consultant @ Migration Observatory)

SAMPLE CASE

Citigroup

In October 2020, the US Office of the Comptroller of the Currency fined Citibank $400 million for deficiencies in risk management, data governance and internal controls. The bank agreed to a consent order and a plan to fix the issues.

Four years later, the Federal Reserve found that Citigroup had made insufficient progress remediating its problems with data quality management. In July 2024, the Fed and the OCC imposed further penalties of about $135.6 million. CEO Jane Fraser pointed to "fragmented tech platforms" and decades of underinvestment in parts of the bank's infrastructure. CFO Mark Mason said the bank had fallen short in data quality management, particularly in regulatory reporting.

Citi invested more than $12 billion in technology in 2023. It has since focused on simplifying its systems: about 2,000 legacy applications were decommissioned over three years, with 130 more retired or replaced in the first quarter of 2025.

Key takeaway: Citi's problem was not a lack of investment but a landscape of many systems with no consistent control of the data between them. Fixing it meant simplifying the systems, not only cleaning the data.

"We fell behind in data, particularly regarding regulatory reporting."

Jane Fraser (CEO @ Citi)

In Short

A data quality framework for many systems does not come from a single tool or a one-off clean-up. It comes from clear definitions, named owners, one golden source per data element and regular checks between systems.

Banks that start with their critical data, measure it against agreed rules and fix errors where they start spend less time correcting reports. The result is figures that management, clients and supervisors can trust.

Frequently Asked Questions

What is a golden source of data?

A golden source is the one system where a data element is created and maintained, such as the client's domicile in core banking. All other systems copy the value from there instead of keeping their own version. When values differ, the golden source wins. Each golden source needs a business owner who is responsible for the quality of the data it holds.

What is the difference between data quality and data governance?

Data governance sets the rules: who owns which data, how it is defined and which system is the golden source. Data quality measures how well the data meets those rules, for example in accuracy, completeness and consistency. Governance without measurement stays on paper. Measurement without governance finds errors but leaves no one responsible for fixing them at the source.

Do we need a new system to improve data quality?

Usually not as a first step. Most improvements come from agreed definitions, clear ownership, automated checks and fixing errors where data is captured. A data quality tool or master data management system can help once these basics are in place. Without them, a new system often becomes one more source of inconsistent data in an already complex landscape.