To check whether your expense categories are right, pick 20 transactions at random from one month, mark each as correct or wrong, and count the wrong ones. If three of twenty are wrong, that is 15%, and it is a signal to fix your rules before you trust a category report. The test takes about ten minutes.
How do I pick 20 rows at random?
Export the month's transactions to a sheet. In a spare column, enter =RAND() next to every row. That gives each row a random number. Copy the column and paste it back as values only, because RAND changes every time the sheet recalculates. Sort by that column and take the first 20 rows. In Google Sheets and Excel the steps are the same. If you prefer a simpler method, take every nth row, where n is the number of rows divided by 20.
What counts as wrong?
A row is wrong if a reasonable person looking at the merchant and amount would put it in a different category. Dispute only the clear misses. A chemist's bill filed under Medical is correct, while a hardware shop under Eating out is not. For borderline cases, such as a supermarket order of household goods that you filed under Groceries, decide your own rule and apply it consistently.
What does a result look like?
| Check | Result | What it means |
|---|---|---|
| Rows sampled | 20 | Enough for a rough read, not a precise one |
| Rows wrong | 3 | Three of 20 is 15% |
| Value of wrong rows | ₹450 + ₹380 + ₹320 = ₹1,150 | Out of ₹14,800 sampled |
| Share of rupees wrong | About 7.8% | 1,150 divided by 14,800 |
| Pattern | All three were small merchant names the rules did not recognise | One alias or rule may fix most |
Two views are useful. By count, 3 of 20 is 15%. By rupees, 1,150 of 14,800 is about 7.8%, since the wrong rows were small. A report grouped by rupees will be less affected than a report that counts transactions. Both numbers tell you something.
Should I trust the reports or fix the rules?
Decide your tolerance before you sample, so that you are not negotiating with yourself afterwards. For example: two or fewer wrong in twenty, carry on; three or more, fix rules first. A sample of twenty is small, so one bad month does not prove the whole history is poor, but it is good evidence. If three wrong rows all share a pattern, such as one type of merchant, a single rule probably cures most of them.
What should I do with the wrong rows?
Correct each one, then ask why it was wrong. If a merchant appears under several names, an alias fixes it. If the rule was too broad, narrow it. In SpendixAI, categorisation happens in stages: your own rules first, then your history and similar merchants, then, only if AI is switched on, an AI model for what is left. Your corrections become history, and you can add merchant aliases. If you prefer, AI can be switched off, and unmatched items are then held for review.
How often should I repeat the test?
Once a quarter is enough, and after any big change such as a new bank account or a new statement. Record the date, the number wrong and the pattern. If the error rate falls from 15% to 5% after you fix rules, you have evidence that the work paid off.
Twenty-row test
- Export one month of transactions and add a RAND column.
- Paste the values, sort, and take the first 20 rows.
- Mark each row correct or wrong.
- Count wrong rows and their rupee value.
- Apply your tolerance and decide: trust, or fix rules first.
- Note the pattern and write one rule or alias to fix it.
Useful ideas start with a clearer picture of your own spending.
Explore your reports

