Use case · Regression
Autonomous AI regression — with confidence intervals you can ship.
Drop a tabular dataset with a numeric target. OctOpus picks the regressor, handles heavy tails, produces calibrated confidence intervals, and deploys a prediction API. Pricing, valuation, propensity, LTV — all the same shape.
TL;DR. Most teams ship point estimates and pretend they're certain. OctOpus produces calibrated intervals (quantile or conformal), reports the empirical coverage on the holdout, and gives you residual diagnostics so you know where the model is wrong.
What OctOpus regresses well
- Pricing — dynamic pricing, market clearing, contract pricing, real-estate valuation.
- LTV / propensity — customer lifetime value, expected revenue, expansion propensity.
- Scoring — credit risk score, quality score, sentiment intensity.
- Forecast as regression — when the time dimension matters, OctOpus escalates to forecasting automatically.
- Scientific / engineering — physical property prediction, material informatics, environmental modeling.
Models the agent rotates through
| Tier | Family | When the agent picks it |
|---|---|---|
| 1 · Baseline (small n) | Ridge / ElasticNet | Datasets under ~500 rows. Strong, interpretable, regularized. |
| 1 · Baseline (mid/large n) | CatBoost / LightGBM | Wide tabular feature sets — usually wins. |
| 2 · Tuned GBM | XGBoost with Optuna · CatBoost with Optuna | When data justifies a search. |
| 3 · Deep tabular | TabPFN / TabNet / FT-Transformer | Complex interactions, mid-size data. |
| 4 · Foundation | TabPFN zero-shot | n < 10k. |
| 5 · Stacking | Linear / GBM stacker over diverse base learners | Decorrelated residuals. |
How OctOpus handles the things regressors usually break on
- Heavy-tailed targets. Log / Box-Cox / Yeo-Johnson transforms, evaluated against the original-scale holdout error.
- Outliers. Robust losses (Huber, quantile), winsorization with explicit reporting.
- Heteroscedasticity. Quantile regression for interval bounds; conformal prediction as a model-agnostic fallback.
- Multi-output. Per-target models with chained features when targets correlate.
- Calibration of intervals. Empirical coverage on the holdout is reported — if the 90% interval covers 85% of the truth, you'll know.
What you get back
- Point prediction plus calibrated lower / upper bound per row.
- R², MAE, MAPE, RMSE — the agent picks the headline per data shape.
- Calibration / coverage report on the intervals.
- Residual plot, QQ plot, error-by-bucket diagnostics.
- SHAP feature importance.
- The
train.py,model.pkl, and a deployable bundle.