Use case · Churn prediction
Autonomous AI churn prediction, calibrated and deployed.
Drop your customer table, get calibrated churn probabilities you can trust — not just rankings. OctOpus picks the model, runs subgroup fairness, validates on a real holdout, and deploys a prediction API your CRM can call.
TL;DR. Most churn models score customers in the right order but lie about the actual probability. OctOpus calibrates by default (isotonic / sigmoid), reports a calibration curve, computes per-subgroup performance, and ships the deploy bundle. The output is something a revenue team can actually act on.
What OctOpus predicts well
- SaaS / subscription churn — month-ahead, quarter-ahead, expansion-vs-contraction.
- Telco churn — contract end-of-life, voluntary cancellation, plan downgrade.
- B2B account risk — health scoring across usage, support, billing, NPS signals.
- E-commerce repeat-purchase risk — RFM-style features auto-engineered from event data.
- Marketplace / two-sided platforms — supply-side and demand-side churn handled separately.
Why calibration matters (and how OctOpus handles it)
A model that says "this customer has a 90% chance of churning" and is right 60% of the time when it says 90% will burn your retention budget. OctOpus runs a calibration step on every classification winner, applies isotonic or sigmoid calibration when the raw scores need it, and reports the calibration curve. The probability is the probability — not just the rank.
Models the agent rotates through
| Tier | Family | When the agent picks it |
|---|---|---|
| 1 · Baseline | CatBoost / LightGBM | Wide tabular customer feature sets — usually wins. |
| 2 · Tuned GBM | XGBoost with Optuna (≥30 trials, TPE, CV objective) | When the data justifies a search. |
| 3 · Deep tabular | TabPFN / TabNet / FT-Transformer | Small-to-mid datasets, complex interactions. |
| 4 · Foundation | TabPFN zero-shot | n < 10k customers; cold-start. |
| 5 · Stacking | Stacking classifier with 3+ diverse base learners | When residuals of different families decorrelate. |
What you get back
- Calibrated churn probability per customer, with confidence band.
- Lift curve and gain chart so revenue ops can size retention plays.
- Per-subgroup performance (segment, plan, region, tenure bucket).
- Top SHAP features per prediction so CSMs can act on the why.
- The
train.pythe agent wrote, fully inspectable. - A deployed prediction API your CRM can call.