liveBook · Manning

Population stability index

As explained in Machine Learning System Design

The population stability index is a divergence measure that quantifies how a population's distribution differs from a reference distribution.

Population stability index (PSI) measures how a prediction population differs from a reference population. In production monitoring, it can serve as an early warning for prediction drift before delayed truth data reveals a model-quality problem. section 14.5.1

A practical monitoring setup treats PSI as one candidate signal rather than an unquestioned decision rule:

  • Track next-day, next-week, and next-month predictions for drift. section 14.5.1
  • Test an alert threshold of PSI greater than 0.2 alongside Wasserstein distance greater than 0.1. section 14.5.1
  • Adjust the reference data for known growth before comparing periods, then refine the approach through historical experimentation. section 14.5.1
  • Use delayed ground-truth metrics to assess model relevance, while retaining drift alerts to detect change earlier. section 14.5.1

How Machine Learning System Design covers it

How deep it goes
Defines an operational threshold and compares PSI with Wasserstein distance, but defers final calibration to historical experimentation.
Examples
Illustrates monthly comparisons using growth-adjusted reference data, including a 15% increase between April 2021 and April 2022.
What you can do after
Operate prediction monitoring by testing PSI alerts alongside Wasserstein distance and validating thresholds against historical data.
How it connects
Connects prediction drift with Wasserstein distance, data drift, delayed ground truth, and model quality.

Also explained in

Compare how 2 books explain population stability index →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Machine Learning System Design. The section links point to the parts of the book it draws on.