Detect Concept Drift in Python with a Simple Statistical Test

Detect concept drift by comparing the mean of recent data against a reference distribution using a z-score-like threshold.

Medium Python 3.9+ Aug 9, 2026 ML engineering pipelines 15 views 0 copies

concept drift statistics ml monitoring z-score

Python code

27 lines

Python 3.9+

import random
import statistics

def detect_drift(recent, reference, threshold=1.5):
    ref_mean = statistics.mean(reference)
    ref_std = statistics.stdev(reference)
    
    recent_mean = statistics.mean(recent)
    drift_score = abs(recent_mean - ref_mean) / (ref_std if ref_std > 0 else 1)
    
    drifted = drift_score > threshold
    return drifted, drift_score

if __name__ == "__main__":
    random.seed(42)
    
    reference_data = [random.gauss(50, 5) for _ in range(100)]
    normal_data = [random.gauss(50, 5) for _ in range(20)]
    drift_data = [random.gauss(70, 5) for _ in range(20)]
    
    print("Normal data drift:")
    drifted, score = detect_drift(normal_data, reference_data)
    print(f"  Score: {score:.3f}, Drift detected: {drifted}")
    
    print("Drifted data detection:")
    drifted, score = detect_drift(drift_data, reference_data)
    print(f"  Score: {score:.3f}, Drift detected: {drifted}")

Output

stdout

Normal data drift:
  Score: 0.327, Drift detected: False
Drifted data detection:
  Score: 3.880, Drift detected: True

How it works

This function computes a drift score as the absolute difference between the recent and reference means divided by the reference standard deviation. A score above the threshold indicates that the recent data is statistically far from the reference baseline, suggesting concept drift. The threshold (default 1.5) is a typical z-score cutoff; you can adjust it based on sensitivity needs. Using statistics.mean and statistics.stdev keeps the implementation dependency-free and works with any iterable of numbers.

Common mistakes

Using population stdev (`pstdev`) instead of sample stdev, which under/overestimates drift for small samples
Not handling zero standard deviation, which causes division by zero
Applying the test when the reference sample is too small, leading to unstable estimates

Variations

Use a sliding window approach where reference data updates over time
Use scipy.stats.ztest for a formal statistical test with p-values

Real-world use cases

Monitoring ML model performance in production by comparing prediction distributions over time.
Triggering retraining pipelines when input feature distributions shift beyond a threshold.
Alerting on drift in user behavior data for recommendation or fraud detection systems.

Detect Concept Drift in Python with a Simple Statistical Test

Python code

Output

How it works

Common mistakes

Variations

Real-world use cases

More from ML engineering pipelines

Tutorials

Quizzes

Python code

Output

How it works

Common mistakes

Variations

Real-world use cases

More from ML engineering pipelines

Keep learning

Tutorials

Quizzes