Outcome feedback into model retraining

Description

The pattern that closes the loop between what happened in the world and what the system learns from it. Includes outcome capture, label propagation back to the prediction record, drift detection, and the retraining cadence. Without this loop, models calcify around their original training data and drift quietly from current reality.

Common failure modes

• 01Outcomes captured but not joined back to the predictions that drove the decision, so the model never sees its own track record
• 02Retraining cadence too slow for the underlying behavior change rate, so the model is always one season behind
• 03Label leakage between training and inference data, producing model performance that looks great in development and disappoints in production
• 04New models retrained without monitoring whether they actually outperform the production model on real traffic, so the new model ships and degrades silently

Tools for this pattern

The capability a stack needs to run this pattern, and the vendors that provide it, on Martech Stack Builder.

• NeedsFeature Store
• RelatedAI & Machine Learning

Recipes composing this pattern