See which risk factors separate a risky customer from a quiet one.

A factor's threshold is the score above which it pushes a customer up a rating.
Crusetra Scoring runs seven factors at every threshold, over the reviews your analysts already decided.

git clone https://github.com/ArslaneSempai-ui/crusetra-scoring cd crusetra-scoring npm ci --ignore-scripts npm run measure:yours -- --customers=your-customers.csv --reviews=your-reviews.csv

The shelving of weights, state 01: Geography catches none of the escalations, and raises no alert.

geography factor: recall 0%kept files: false alerts 0%

finding 01no weight here

Every file in our test set lives in an EU country, and no high-risk country list names any of them, because no official country list was copied into the data. So the geography factor scores 0% on both sides: the factor a vendor sells hardest carries no weight here, by design.

0%catches 0% of escalated files with the geography factor at 0.50, interval 0% to 8%, on 42 files

0%false alerts 0% at 0.50, interval 0% to 8%, on 42 files kept at their rating

The shelving of weights, state 02: The product factor catches 92.9% of escalated files, and alerts on 23.8% of the rest.

product factor: recall 92.9%kept files: false alerts 23.8%

finding 0223.8% of quiet files alerted

Scoring only the products a client holds catches 92.9% of the escalated files at threshold 0.50, because correspondent accounts, private banking and crypto rails carry the signal. The same factor alerts on 23.8% of the files kept at their rating, and each alert is a review someone has to run.

92.9%catches 92.9% of escalated files with the product factor at 0.50, interval 81% to 98%, on 42 files

23.8%false alerts 23.8% at 0.50, interval 13% to 39%, on 42 files kept at their rating

The shelving of weights, state 03: No factor is sure of catching 90% of the escalated files.

recall floor 90%: no factor reaches itbehaviour: 51.6%, no false alert

finding 03closest 81%

The tool keeps only the settings whose confidence interval stays above 90% of true files, then among those takes the fewest false alerts. We chose that 90% floor ourselves, and you can move it. No factor reaches it on the 42 files: product comes closest at 81%, and behaviour, which raises no false alert, at 51.6%.

81%product at 0.50 comes closest: 81%, interval 81% to 98%

51.6%behaviour at 0.50: 51.6%, with no false alert

The shelving of weights, state 04: The product factor catches 79.4% of generated files against 92.9% of the written ones.

written files: recall 92.9%generated files: recall 79.4%

finding 0479.4% against 92.9%

The same setting catches 92.9% of the files we wrote and 79.4% of the ones a script generated. Here the generated set is harder, and it is still counted on its own, so the rates above come from the files we wrote.

92.9%catches 92.9% with product at 0.50, on the files we wrote, interval 81% to 98%

79.4%catches 79.4% at the same setting, on 126 files a script generated, interval 71% to 86%

The shelving of weights, state 05: Run the same measurement on your own review outcomes.

84 files written252 files generated

finding 05one run, one report

The rates here come from our public test set: 84 files written by hand, 42 files escalated and 42 files kept at their rating, with 252 files generated from them. Give the tool two CSV files from your systems, your review outcomes and the customer attributes, and it writes the report into the folder they are in.

42 files84 files written, 42 files escalated and 42 files kept at their rating, seven per typology

126 files252 files generated, 126 files escalated and 126 files kept at their rating, all kept separate

Try the Scoring instrument on our public test set.

Crusetra · ScoringOpen the live instrument Each risk factor at each threshold: the risky customers it catches, and the quiet ones it flags. Each quiet customer flagged costs a reviewer the time of an extra review. Live from our public test set.

See pricing

Measured on 42 escalated files and 42 files kept at their rating, all written by us, at every threshold, and the marked cell, product at 0.50, comes closest to the 90% floor we set, which none reaches. Generated files are counted on their own, and the live instrument shows them.

Crusetra, explained.

The five Scoring findings

also on YouTube