Models and experiments
Decision models and machine learning, measured on your data.
Grayson, Finic's decision model, answers questions about each request in plain language and works from day one. Logistic regression and gradient-boosted trees learn your own outcomes. Experiments tell you which mix wins.
VariantROC-AUC
- Grayson alone0.912
- Logistic regression · fields0.871
- XGBoost · fields0.894
- Logistic regression · fields + GraysonWinner0.949
- XGBoost · fields + Grayson0.941
Gain vs. live+0.02895% interval+0.015 to +0.041Recall at block71%
Models and experiments
What you get
Works before you have labelsGrayson needs no training data, so you catch fraud before you have the chargebacks a model learns from.
Reads what trees can'tDescriptions, emails, names, addresses, notes and item titles become signals.
Learns your outcomesLogistic regression and XGBoost-compatible trees train in Kairos on your numbers and Grayson's answers, or you upload your own.
Honest resultsVariants train on older data and are tested on newer data, with ROC-AUC, PR-AUC, a 95% interval on the gain, and precision and recall at your thresholds.
How an experiment runs
- 01Propose variantsPick the models and inputs to compare.
- 02Split by timeTrain on the oldest labeled data, choose on the next, test on the newest.
- 03Wait for labelsA request counts as labeled only once its outcome window has passed, so late fraud doesn't train as legitimate.
- 04Apply the winnerAs a draft for review, or as the workflow's next version.