For ML & data teams
Check your evaluation workflow with controlled rare events
Test splitting, metrics and ingestion with a reproducible rare-label fixture. Inspect 10,000 transactions with explicit channel-dependent probabilities before adapting the scenario.
Free preview. Sign in to save an editable copy. Normal credits apply when you export.
Known rules. Measured counts. Repeatable checks.
- 1
Choose a controlled scenario
Start with 10,000 transactions and a boolean fraud label: 2% probability for card-present transactions, 4% online. The published example reports the actual counts.
- 2
Inspect and vary the conditions
Review the preview and conditional rules in Ground. Change the seed or probabilities to exercise class imbalance and missing-value handling.
- 3
Check the evaluation code
Export the fixture to test stratified splits, confusion matrices and ingestion. For model selection, evaluate separately on a representative real holdout.
A look at the actual output
All files and measured counts →| timestamp | amount_usd | merchant_category | channel | state | is_fraud |
|---|---|---|---|---|---|
| 2025-04-09T17:36:34Z | 344.96 | grocery | card_present | MI | false |
| 2025-02-21T00:07:47Z | 1978.7 | electronics | online | CA | true |
| 2025-06-19T12:14:40Z | 962.87 | other | card_present | PA | false |
| 2025-03-13T18:53:16Z | 999.29 | online_retail | card_present | NC | false |
| 2025-06-01T09:39:15Z | 256 | online_retail | online | NY | false |
What to check
Confirm boolean parsing, count positives by channel and check that your split and metrics handle a small positive class. Finite samples do not match probability quotas exactly.
What this example can tell you
These labels follow a specified simulation, not verified fraud behavior. There are no card histories or explicit amount-to-label relationship. This fixture cannot establish real model accuracy or improved utility.