Exporting historical data for targeted model optimization

In production, enabling historical data collection for an Adaptive Model is a targeted decision, not a default operating mode. This topic explains the specific purpose of the Record historical data option, when to enable it for a model, and how to manage scale with sampling and retention settings. You learn how to use historical model data for offline analysis, challenger testing, and deeper diagnostics, while avoiding unnecessary storage growth.

Welcome to this session on exporting historical data in Pega Customer Decision Hub. This session addresses one question: When should you turn on historical data collection for a model, and how do you use it to improve that model?

Adaptive models in Customer Decision Hub are already smart and self-learning. They improve continuously from customer outcomes. In day-to-day operations, you do not need to collect every historical record for every model all the time. Turn historical collection on for a specific model when you want deeper insight, stronger evidence, and better optimization decisions.

Where does this help most? First, diagnostics. If a model is underperforming, historical data lets you inspect the real predictor values, context, and outcomes behind its behavior. You can identify missing fields, weak signals, or noisy inputs that make learning less effective.

Second, challenge your testing. You can use historical records to test alternative model settings and compare results before making production changes. That means lower risk, cleaner experiments, and better confidence when you promote improvements.

Third, external analysis. You can export the data and run deeper exploratory analysis, including advanced reporting and visual analysis. This is especially useful when you want to compare patterns across channels, actions, and customer segments.

The biggest practical issue is scale. Historical data can grow very fast. Sampling means you choose how much positive and negative outcome data to keep. Response data is usually imbalanced. For example, clicks or accepts are often much rarer than impressions or rejects. A common setup is to keep positive outcomes at a high sample rate and reduce the negative sample rate to keep files manageable. You preserve the valuable learning signal without accumulating excess volume.

What exactly is in the exported data? You get model context, predictors, decision-related fields, and monitoring data. That means you can reconstruct what decision happened for which customer context, with what outcome, and what the model had learned at that moment. That combination is highly valuable for model optimization.

To wrap up, do not treat historical collection as always-on telemetry. Treat it as a targeted performance instrument. Use it when you need evidence. Scope it tightly. Control volume. Extract insight. Improve the model. Then move to the next high-impact opportunity.

Thank you for watching!