Help me pick a dataset for churn prediction with recent labels, low missingness, and commercial rights

Here’s a practical way to choose a churn dataset that fits your constraints: recent labels, low missingness, and commercial rights.

Best-fit public options to consider

1) Telco Customer Churn

• Why it fits: Very common churn benchmark, relatively clean, easy to use.
• Data quality: Low to moderate missingness; usually manageable.
• Recent labels: Not truly “recent” in a live-business sense, but labels are straightforward and usable.
• Commercial rights: Depends on the source/license. You need to verify the dataset license carefully before commercial use.
• Good for: Fast prototyping and baseline churn modeling.

2) KKBox Churn Prediction Dataset

• Why it fits: Real-world subscription churn from a music streaming service.
• Data quality: Generally decent; some tables may require preprocessing, but missingness is not typically the main issue.
• Recent labels: Better than many older churn datasets because it reflects subscription behavior over time.
• Commercial rights: Often available through competition/research platforms, but commercial usage may be restricted. Check the exact terms.
• Good for: More realistic churn modeling with event/history data.

3) Online retail / subscription SaaS datasets from Kaggle or GitHub

• Why it fits: Sometimes have recent timestamps and lower missingness.
• Data quality: Varies a lot.
• Commercial rights: Often not safe for commercial use unless explicitly licensed for it.
• Good for: Exploration only, unless licensing is very clear.

What to prioritize

If your requirements are strict, use this filter:

1. License allows commercial use

   • Look for CC0, CC-BY 4.0, or a clearly commercial-friendly license.
   • Avoid datasets with “research only,” “non-commercial,” or unclear provenance.

2. Recent labels

   • Prefer datasets with labels from the last 1–3 years if you need current behavior.
   • Check whether churn is defined by a recent observation window and follow-up period.

3. Low missingness

   • Aim for <5–10% missingness in key features.
   • Avoid datasets where churn labels are present but core predictors are heavily incomplete.

My recommendation

If you need something usable quickly:

• Start with Telco Customer Churn for modeling workflow validation.
• But only use it commercially if the source license explicitly permits it.

If you need stronger realism:

• Consider KKBox, but verify whether the competition/data terms permit commercial use.

If commercial use is non-negotiable:

• You may be better off using a commercially licensed provider or building your own churn dataset from CRM/product logs, because many public churn datasets have licensing ambiguity.

Quick checklist before choosing

• Is the license explicit and commercial-friendly?
• Are the labels recent enough for your business case?
• Is missingness low in the main predictive features?
• Are there any privacy/PII restrictions?
• Does the dataset match your churn definition?

If you want, I can give you a shortlist of specific datasets with license notes or help you find one by industry (telecom, SaaS, retail, fintech, etc.).