Live dashboard
See recent volume, products, topics, and signal counts.
Open dashboard →Built from the full CFPB archive
See what people report.
Find what changed.
ConductWatch turns public banking complaints into simple, explainable signals. It helps people decide what to read next; it does not decide who is at fault.
17M complaints4 signals to review71.4% macro-F1
Real signal map
What changed most?
Review
Unable to get your credit report or credit scorePayday loan, title loan, personal loan, or advance loan
- Observed · Jul 13, 2026–Jul 19, 2026
- 7
- Expected · median, Jan 12, 2026–Jul 12, 2026
- 0
- Vs. average · Jun 15, 2026–Jul 12, 2026
- +675%
- Signal score · unusual distance
- 7.00
Size = complaints in the observed week. Higher = further above its own norm. Right = faster than the prior four weeks.
Plain meaning: “Expected” is the median weekly count across the previous 26 complete weeks. “Matching evidence” is a recent public narrative with the same product and issue; it is context, not proof.
Real state borders · interactive depth map
Where complaint reporting concentrates
Color and raised-state motion show relative concentration. This uses geographic state boundaries; Alaska and Hawaii appear as map insets. Counts describe submitted reports, not proven harm.
Why it exists
Millions of records.
A clearer place to start.
Complaint data can reveal changing customer problems, but a raw spreadsheet is difficult to scan.
ConductWatch finds unusual weekly movement, groups common language, and keeps the source evidence close to every result.
Four working tools
Move from pattern to evidence.
No account is required. Public data is ready to explore, and your own CSV stays in your browser.
4
Signal review
See which product-and-issue pairs moved beyond their own history.
Review signals →96% top-three
Model lab
Try the local triage helper and read the full scorecard.
Open model lab →CSV
Your own data
Check volume and data quality without uploading to a server.
Analyze a file →Measured honestly
A score needs
an explanation.
71.4Macro-F1
all classes count equally
all classes count equally
71.8Accuracy
27,000 later complaints
27,000 later complaints
The model was tested on later complaints it did not train on. It suggests queues; it does not make regulatory findings.
Read the model card →Open and inspectable
Start with a change.
End with the evidence.
Every public chart comes from the full, deduplicated archive.
Enter ConductWatch →