On 17 June 2012, the IT team at the RBS group upgraded the software that processed overnight updates to customer accounts. Testing financial software was the step that fell short. The UK Financial Conduct Authority reported on 20 November 2014 that over 6.5 million customers at RBS, NatWest and Ulster Bank were affected for several weeks.
Many could not use online banking or see accurate balances at cash machines. Some missed mortgage payments, and some companies could not meet payroll. The banks also applied the wrong interest to accounts and sent out inaccurate statements. The FCA fined the three banks £42 million, and the Prudential Regulation Authority added £14 million. The FCA found "inadequate testing procedures for managing changes to software".
The techniques are the same, the boundaries move
You test a banking or payments system with the same techniques you use anywhere else. Equivalence partitioning splits inputs into groups that the system should treat the same way. Boundary value analysis tests the values at the edges of those groups, where defects tend to cluster. Both are in the ISTQB Foundation Level syllabus.
In finance the boundaries of those groups may sit at half a cent, at midnight on the last day of the month, or at the moment someone edits a record. Some are written into law with a clause number. A tester that only checks the input limits shown on the screen may never test the rounding, cut-off or audit rules.
Rounding and decimal precision
Most programming languages store numbers like 0.1 as binary floating point. That format cannot hold most decimal fractions exactly. The Python documentation shows that 0.1 + 0.1 + 0.1 == 0.3 returns False. The notes on round() add that round(2.675, 2) gives 2.67 instead of the expected 2.68. A developer that stores money as a float can lose or gain a cent without any error message.
Some regulations state which way a cent should go. When the euro was introduced, Council Regulation (EC) No 1103/97 of 17 June 1997 set the rule in Article 5. It covers an amount to be paid or entered in an account after it is converted into euros under Article 4. That amount is rounded to the nearest cent, and a result that is exactly half-way is rounded up. Article 4 says the conversion rates have six significant figures and must not be rounded or cut short during a conversion.
Your test cases come straight from that rule. Take an amount that converts to 2.67 and a fraction of a cent.
In the Python example above, the middle row exposes the mismatch: round(2.675, 2) gives 2.67, but the requirement expects 2.68. Some systems use a different rule, called banker's rounding, where an exact half goes to the nearest even cent. Under that rule 2.665 becomes 2.66. Neither rule is wrong in general. But the requirement has to name one, and your expected result has to match it.
Then test the totals. Split $10.00 three ways and each share rounds to $3.33, which adds up to $9.99. Where does the missing cent go? Is it added to the first share, the last share, or a separate account? A test that only checks each line can pass while the ledger is a cent short every time.
Dates, cut-offs and the overnight batch
Batch processing means the system runs a set of jobs together, often overnight, in a fixed order. Suppose one job posts the day's payments, and the next job calculates interest on the new balances. If the first job fails or runs late, every job after it may work on the wrong numbers. The FCA did not give a job-by-job cause for RBS. It said the bank found problems with the upgrade and removed it without first testing what that would do. The upgraded software was not compatible with the old version.
A date defect may not show until the system reaches that date. On 1 January 2010, about 30 million chip and PIN cards in Germany stopped working, The Irish Times reported on 7 January 2010. The chips could not handle the change to the year 2010. The card maker Gemalto said it was to blame. The Register reported the same week that customers could not use the cards at cash machines or shop tills.
So what are the partitions for a date? Take a bank with a 17:00 cut-off for same-day payments. These are the classes and boundaries worth a test case:
- Time of day: 16:59:59, 17:00:00 and 17:00:01. The requirement should say which side 17:00:00 falls on. If it does not, that is a defect in the requirement, and you raise it before you write the test.
- Type of day: a business day, a weekend and a public holiday. A payment sent on Saturday may need to post on Monday.
- End of month: months with 28, 29, 30 and 31 days. Monthly interest and loan due dates can fall on the 31st, or on 29 February in a leap year.
- Year change: 31 December and 1 January, in every time zone the system serves.
Then test the batch itself as a system. Run the interest job twice on the same day and check that interest posts once. Skip a night and check that the next run catches up. Stop a job halfway, restart it, and compare the balances with a clean run. These tests take time to set up. Also test the rollback: remove an upgrade in a test environment and compare the balances with a clean run.
Audit trails: can you rebuild the original?
An audit trail is a record of every change to the data: who changed it, when, and what the old value was. In finance it is often a legal requirement. On 12 October 2022, the US Securities and Exchange Commission adopted changes to Rule 17a-4 for broker-dealers. Firms may now keep electronic records in a way that "permits the recreation of an original record if it is altered, over-written, or erased".
That phrase gives you a test oracle, which is the rule that tells you whether a result is right. Change a customer's payee details, then try to rebuild the record as it was before the change. Delete a transaction and check that the trail still shows it existed. Log in with an admin account and try to edit the trail itself. Compare the timestamps across systems that may sit in different time zones. If you cannot rebuild the original from the trail, the test fails, even when every screen looks correct.
Testing financial software against a written rule
When a requirement in finance is copied from regulation, the rule often states its own boundaries. Take the annual percentage rate (APR) that US lenders must disclose. Picture a regular loan with a set end date, such as a car loan with one advance and equal monthly payments. For that loan, Regulation Z section 1026.22 says a disclosed APR is accurate if it is within 1/8 of 1 percentage point of the correct rate.
Suppose the correct APR for a regular loan is 6.000%. The accurate partition runs from 5.875% to 6.125%. If rates are stored to 0.001 percentage point, your boundary tests are 5.874%, 5.875%, 6.125% and 6.126%.
Put the clause number in each test case. When the rule changes, you can find every test that depends on it. In many banks a compliance team, not the tester, may decide what the regulation means for the product. Your job is to show that the software does what that reading says, at every boundary.
A tester that works this way may find defects before a regulator does. If you want to practise boundary value analysis and equivalence partitioning on cases like these, Nopmark's ISTQB CTFL certification training includes practice exams and hands-on labs to build them.
Sources
- FCA fines RBS, NatWest and Ulster Bank Ltd £42 million for IT failures, Financial Conduct Authority, 20 November 2014
- Council Regulation (EC) No 1103/97, Articles 4 and 5, adopted 17 June 1997; published in Official Journal of the European Communities L 162, 19 June 1997
- Floating-Point Arithmetic: Issues and Limitations, Python Software Foundation, current docs accessed 2 October 2026
- Built-in Functions: round(), Python Software Foundation, current docs accessed 2 October 2026
- Software bug leaves Germans cash-strapped, The Irish Times, 7 January 2010
- Year 2010 bug wreaks havoc on German payment cards, The Register, Dan Goodin, 6 January 2010
- SEC Adopts Amendments to Electronic Recordkeeping Requirements, US Securities and Exchange Commission, 12 October 2022
- Regulation Z, 12 CFR 1026.22 Determination of annual percentage rate, Consumer Financial Protection Bureau, accessed 2 October 2026