📊 Classic MLID: psi-data-drift

Data Drift & PSI

Q: Write $PSI = \sum_i (A_i - B_i)\ln\frac{A_i}{B_i}$ and define $A_i$/$B_i$; why is PSI = 0 when $A_i = B_i$? How does it relate to KL divergence?

【Data Drift & PSI 核心考点 1】Write $PSI = \sum_i (A_i - B_i)\ln\frac{A_i}{B_i}$ and define $A_i$/$B_i$; why is PSI = 0 when $A_i = B_i$? How does it relate to KL divergence?。技术解析：Definition: PSI (Population Stability Index) is the most common metric for monitoring input-distribution drift, quantifying the binned divergence between the baseline (e.g. training) distribution and the current one: PSI = \sum_i (A_i - B_i)\ln\frac{A_i}{B_i}, where A_i is the sh... 完整推导与工程落地见：https://talentme.airsota.com/resources/tech/roadmap/mindmap/ML/card/psi-data-drift

Q: PSI thresholds 0.25 and the action for each; how to handle zero bins ($B_i = 0$ or $A_i = 0$)?

【Data Drift & PSI 核心考点 2】PSI thresholds 0.25 and the action for each; how to handle zero bins ($B_i = 0$ or $A_i = 0$)?。技术解析：Definition: PSI (Population Stability Index) is the most common metric for monitoring input-distribution drift, quantifying the binned divergence between the baseline (e.g. training) distribution and the current one: PSI = \sum_i (A_i - B_i)\ln\frac{A_i}{B_i}, where A_i is the sh... 完整推导与工程落地见：https://talentme.airsota.com/resources/tech/roadmap/mindmap/ML/card/psi-data-drift

Q: Covariate drift vs concept drift: definitions, one real example each, and how each affects the model differently?

【Data Drift & PSI 核心考点 3】Covariate drift vs concept drift: definitions, one real example each, and how each affects the model differently?。技术解析：Definition: PSI (Population Stability Index) is the most common metric for monitoring input-distribution drift, quantifying the binned divergence between the baseline (e.g. training) distribution and the current one: PSI = \sum_i (A_i - B_i)\ln\frac{A_i}{B_i}, where A_i is the sh... 完整推导与工程落地见：https://talentme.airsota.com/resources/tech/roadmap/mindmap/ML/card/psi-data-drift

Q: Why monitor PSI instead of online AUC alone? What are PSI's limitations (e.g. sensitivity to binning)?

【Data Drift & PSI 核心考点 4】Why monitor PSI instead of online AUC alone? What are PSI's limitations (e.g. sensitivity to binning)?。技术解析：Definition: PSI (Population Stability Index) is the most common metric for monitoring input-distribution drift, quantifying the binned divergence between the baseline (e.g. training) distribution and the current one: PSI = \sum_i (A_i - B_i)\ln\frac{A_i}{B_i}, where A_i is the sh... 完整推导与工程落地见：https://talentme.airsota.com/resources/tech/roadmap/mindmap/ML/card/psi-data-drift

Q: The action pipeline when PSI ≥ 0.25 (attribute → retrain → rollback); how to distinguish real drift from data leakage or definition changes (feature/口径 change)?

【Data Drift & PSI 核心考点 5】The action pipeline when PSI ≥ 0.25 (attribute → retrain → rollback); how to distinguish real drift from data leakage or definition changes (feature/口径 change)?。技术解析：Definition: PSI (Population Stability Index) is the most common metric for monitoring input-distribution drift, quantifying the binned divergence between the baseline (e.g. training) distribution and the current one: PSI = \sum_i (A_i - B_i)\ln\frac{A_i}{B_i}, where A_i is the sh... 完整推导与工程落地见：https://talentme.airsota.com/resources/tech/roadmap/mindmap/ML/card/psi-data-drift

数据漂移与 PSI

🎯Core Definition

PSI (Population Stability Index) is the most common metric for monitoring input-distribution drift, quantifying the binned divergence between the baseline (e.g. training) distribution and the current one:

PSI = \sum_i (A_i - B_i)\ln\frac{A_i}{B_i}

, where

A_i

is the share of actual samples falling in bin

i

and

B_i

the expected share; bins are typically deciles of the baseline distribution, empty actual bins are merged with neighbors to avoid division by zero, and PSI is essentially a symmetrized KL divergence. Rule of thumb:

PSI < 0.1

stable;

0.1 \le PSI < 0.25

needs investigation;

PSI \ge 0.25

signals significant drift and triggers retraining. Two drift types: covariate drift (

P(X)

changed, e.g. a new customer mix) and concept drift (

P(Y|X)

changed, e.g. default rates rose during COVID).

💡Use Cases

a must-have on weekly/monthly monitoring dashboards for credit and anti-fraud models; interview questions like "how do you detect model degradation in production" or "what to do when PSI exceeds the threshold"; compute per-feature PSI to rank offending variables, or one overall PSI over the model score, then localize root causes by PSI ranking.

⚡Key Problems Solved

unlike watching only online metrics such as AUC (labels arrive late, and no localization of the offending feature), PSI detects input-distribution drift per feature ahead of labels; combined with tiered thresholds (

<0.1

stable /

0.1\text{-}0.25

watch /

>0.25

drift) and retraining triggers, it closes the monitor—diagnose—retrain loop.

🎯5 High-Frequency Exam Points

Write

PSI = \sum_i (A_i - B_i)\ln\frac{A_i}{B_i}

and define

A_i

B_i

; why is PSI = 0 when

A_i = B_i

? How does it relate to KL divergence?

PSI thresholds <0.1 / 0.1–0.25 / >0.25 and the action for each; how to handle zero bins (

B_i = 0

A_i = 0

Covariate drift vs concept drift: definitions, one real example each, and how each affects the model differently?

Why monitor PSI instead of online AUC alone? What are PSI's limitations (e.g. sensitivity to binning)?

The action pipeline when PSI ≥ 0.25 (attribute → retrain → rollback); how to distinguish real drift from data leakage or definition changes (feature/口径 change)?

📖 In-depth Guide:📄 ml-math-and-eval-metrics →

Updated 2026-08-12

🎯

Test Your Knowledge: Practice Questions for "Data Drift & PSI"

Single choice pitfall questions with instant feedback and mistake tracking.

🚀 Start Card Practice ➔

← Previous CardSHAP & LIME Next Card →Learning to Rank

🔗 More Classic ML Knowledge Cards

AdaBoost Derivation Bagging & Random Forest Baum-Welch (HMM EM)GBDT Negative Gradient