Drift Detection

The Drift Problem in Production

AI systems perform beautifully in testing, then degrade in production. Not because the code is buggy, but because the world has changed. Data has shifted. Distributions have diverged. The gap between training data and production data grows larger every day.

Critical statistic: VentureBeat (2019) surveyed 1,800 ML teams and found that 91% of ML models degrade within 9 months of deployment. Not all fail completely, but performance drifts downward.

The solution is continuous monitoring for drift and automated response protocols. This guide covers the full toolkit: what drift is, how to detect it, and how to act on signals before users notice.

Types of Drift: A Taxonomy

Data Drift (Covariate Shift)

Definition: The distribution of input features P(X) has changed, but the relationship between features and labels P(Y|X) has not.

Example: You trained a credit scoring model on data from 2020-2022. In 2024, post-pandemic economy changes the distribution of applicant income, employment sector, and debt levels. Your model sees different X values than it saw in training. But your decision boundaries (P(Y|X)) are still valid—high income still predicts lower default risk.

Detection method: Compare feature distributions between training (reference) data and recent production data. Use KS test, Chi-squared test, or PSI.

Impact: Moderate to high. The model still makes correct decisions given the input, but it's being asked questions outside its training distribution. Edge cases become common.

Concept Drift

Definition: The relationship between features and labels has changed. P(Y|X) has shifted. The same input features now have different predictive meaning.

Example: You trained a spam classifier on 2023 emails. In 2024, spammers adjust tactics. What previously indicated spam (certain keywords, sender patterns) no longer does, or new patterns emerge. The underlying concept of "what is spam" has evolved.

Detection method: Track model performance on recent predictions. If accuracy drops despite inputs looking similar to training data, suspect concept drift. Ground truth comparison is needed.

Impact: Very high. The model's decision boundaries are now wrong. Requires retraining.

Label Shift

Definition: The distribution of labels P(Y) has changed, but the conditional probability P(X|Y) has not.

Example: In your training data, 30% of credit applicants defaulted. In production, economic conditions improve and only 15% default. The input features given a default (what defaults look like) haven't changed, but defaults are rarer.

Detection method: Track the class distribution in predicted labels. If baseline was 30% positive but you're now predicting 15%, label shift may have occurred. Check via stratified sampling for ground truth.

Impact: Low to moderate. Model decisions are still reasonable; it's just that positive cases are rarer. May need probability calibration.

Feature Drift

Definition: One or more input features have changed in a way that breaks their meaning or range.

Example: A feature was "annual_income_USD". In production, a data pipeline change now includes bonus and stock options. The feature still called "annual_income_USD", but now means something different. Or the currency changes. Or the source system switches to a new API that has different data quality.

Detection method: Monitor feature distributions and anomalies. Check min, max, mean, median for each feature over time. Alert if values go outside expected range.

Impact: Can be very high if the feature change is structural. Model makes nonsensical decisions on corrupted features.

Upstream Drift

Definition: Dependencies (data pipelines, feature stores, upstream models) that feed into your model have changed.

Example: Your customer churn model depends on features from an upstream recommendation system (e.g., "user_engagement_score"). That recommendation system gets retrained with new data. Its outputs change. Your model now receives different feature values even though its own training data and code haven't changed.

Detection method: Monitor not just your model, but all upstream systems. Test the end-to-end pipeline monthly. Alert if intermediate outputs change.

Impact: High if upstream system provides critical features. Hard to diagnose because changes appear subtle.

Statistical Tests for Drift Detection

Kolmogorov-Smirnov (KS) Test for Continuous Features

What it does: Tests whether two continuous distributions are different.

Null hypothesis: The distribution of feature X in training data is the same as in recent production data.

Test statistic: KS = max|F_train(x) - F_prod(x)| where F is the cumulative distribution function.

Interpretation: If p-value < 0.05, reject the null. The distribution has likely drifted.

Limitations: Sensitive to outliers. Requires sufficient sample size in both distributions (~100+ observations).

Chi-Squared Test for Categorical Features

What it does: Tests whether the distribution of a categorical feature has changed.

Null hypothesis: The proportions of each category are the same in training and production.

Test statistic: Chi-squared = Σ (observed - expected)² / expected

Interpretation: High chi-squared value → significant difference. p-value < 0.05 suggests drift.

Population Stability Index (PSI) - Detailed Section Below

See dedicated section on PSI, which is the most practical choice for most drift detection tasks.

Population Stability Index (PSI) Deep Dive

What PSI Measures

PSI quantifies how much a feature distribution has shifted between two periods. It's useful for both continuous and categorical features. Unlike statistical tests (which give p-values), PSI gives an interpretable magnitude of drift.

PSI Formula

PSI = Σ (% in production - % in baseline) × ln(% in production / % in baseline)

For continuous features, bin the data first:
1. Create 10 equal-frequency bins using baseline distribution
2. Count % of observations in each bin for baseline period
3. Count % of observations in each bin for current production period
4. Apply formula

Interpretation Thresholds

PSI Value	Interpretation	Action
< 0.1	No significant population change	No action. Normal variation.
0.1 - 0.25	Small population change	Monitor closely. No immediate action, but watch trend.
0.25 - 1.0	Moderate population change	Investigate. Understand what changed. May need retraining soon.
> 1.0	Large population change	Strong signal of drift. Immediate investigation and likely retraining needed.

PSI Worked Example: Credit Scoring

Baseline (training data, 2022-2023): 50K applicants. Income distribution:

• Bin 1 ($0-30K): 15%
• Bin 2 ($30-50K): 25%
• Bin 3 ($50-75K): 30%
• Bin 4 ($75-100K): 20%
• Bin 5 ($100K+): 10%

Recent production (2024): 50K recent applicants. Income distribution:

• Bin 1: 8%
• Bin 2: 20%
• Bin 3: 35%
• Bin 4: 28%
• Bin 5: 9%

PSI calculation (per bin):

Bin	Baseline %	Recent %	Difference	Ratio	ln(Ratio)	Contribution
1	15%	8%	-7%	0.533	-0.630	0.044
2	25%	20%	-5%	0.800	-0.223	0.011
3	30%	35%	+5%	1.167	0.154	0.008
4	20%	28%	+8%	1.400	0.336	0.027
5	10%	9%	-1%	0.900	-0.105	-0.001

Total PSI = 0.044 + 0.011 + 0.008 + 0.027 - 0.001 = 0.089

Interpretation: PSI = 0.089 < 0.1, indicating minimal drift. The income distribution has shifted slightly (fewer low-income applicants, more mid-to-high), but it's not a major change. Continue monitoring monthly.

CUSUM and EWMA for Continuous Monitoring

These methods (covered in detail in rater drift detection) also apply to model output drift.

CUSUM for model performance: Track cumulative deviation of model accuracy from baseline. If CUSUM breaches threshold, accuracy has degraded.

EWMA for model output drift: Track exponentially weighted moving average of model confidence or predicted probability. Sudden drops in confidence (even if accuracy hasn't formally degraded) may signal concept drift.

Windowed Monitoring: Fixed vs. Sliding Windows

Fixed Window Approach

Idea: Compare training period (fixed baseline) to current period (fixed evaluation window). E.g., compare 2022-2023 data (baseline) to May 2024 data (current).

Advantage: Simple. Clear reference point.

Disadvantage: Can miss slowly creeping drift. If drift happens gradually over 12 months, you only detect it in month 12.

Sliding Window Approach

Idea: Use a rolling baseline. Compare recent data (e.g., past 2 weeks) to older data (e.g., 2 weeks before that). Continuously update the baseline.

Advantage: Detects gradual drift faster. More sensitive to recent changes.

Disadvantage: Can be noisy if recent data has natural fluctuation. May flag false positives.

Window Size Selection

How big should each window be? Tradeoff between statistical stability and timeliness:

• Very small (1 day): Fast detection but noisy. May alert on normal fluctuation.
• Small (1 week): Good balance. Detects drift within days of occurrence.
• Medium (1 month): More stable. Less noise. Slower detection (1-2 month lag).
• Large (1 quarter): Very stable. May miss drift until it's severe.

Recommendation for most ML systems: Sliding window with 1-week windows for high-criticality models, 2-week for medium, 1-month for low-risk models.

Embedding Drift: Semantic Shift Detection

The Problem with Feature-Level Drift Detection

For text-based models, traditional feature drift detection fails. Features might appear stable (word frequency distributions identical), but semantic meaning has shifted. E.g., the word "virus" meant one thing in 2019, another in 2020.

Embedding Distance Metrics

Approach: Use a pre-trained embedding model (e.g., sentence-BERT) to convert text into embeddings. Compare embedding distributions between baseline and recent data using distance metrics.

Metrics:

• Cosine distance: Average cosine distance between baseline embeddings and recent embeddings. If distance > threshold, semantic drift likely.
• Wasserstein distance: Earth mover distance between embedding distributions. Captures overall shift in embedding space.
• Maximum Mean Discrepancy (MMD): Statistical test comparing embedding distributions. More principled than naive distance measures.

Worked Example

Task: Detect sentiment drift in customer reviews. Baseline: Q1 2024 reviews. Recent: Q2 2024 reviews.

Method: Embed 1000 baseline reviews and 1000 recent reviews using sentence-BERT. Calculate pairwise cosine distances between each recent embedding and nearest baseline embedding.

Result: Average nearest-neighbor distance = 0.15 (on 0-1 scale). Threshold is 0.10. Distance > threshold → semantic drift detected.

Investigation: Review the reviews with highest distances. Find that Q2 reviews discuss new product features (not discussed in Q1), causing semantic shift.

Action: Retrain sentiment model on data including Q2 reviews.

Multivariate Drift Detection

Why Univariate Detection Isn't Enough

You can monitor each feature individually for drift. But drift can occur in the joint distribution without showing up in any individual feature. E.g., income and debt are correlated. Each individually stable, but correlation structure changes.

Mahalanobis Distance

Formula: D = √((x - μ)ᵀ Σ⁻¹ (x - μ))

Interpretation: Generalized distance accounting for feature correlation and variance. Points far from the training distribution have high Mahalanobis distance.

Monitoring: For each recent observation, calculate Mahalanobis distance from training mean. If >95th percentile distance from training, flag as anomalous.

PCA-Based Approach

Idea: Project data into principal component space. Compare PC distributions between training and recent data.

Advantage: Reduces dimensionality. Makes visualization easier.

Disadvantage: Depends on which PCs are retained. Loss of information in dropped components.

Acting on Drift Detection

The Four Response Modes

1. Monitor More Closely - PSI = 0.15-0.25. Small drift detected, but not actionable yet. Increase monitoring frequency from monthly to weekly. Set up alerts for escalation.

2. Retrain - PSI > 0.25 or accuracy degradation detected. Model needs fresh data. Collect recent ground truth, retrain on baseline + recent, validate on holdout, push to production.

3. Rollback - Model performance dropped unexpectedly (not due to feature drift, but something went wrong). Roll back to previous model version. Investigate what happened. Consider downtime preferable to bad predictions.

4. Escalate to Review - Drift detected but cause unclear. Unusual patterns in features (e.g., feature missing >10% of values). Upstream model changed. Human review needed before making retraining or rollback decision.

Decision Tree for Response

DRIFT DETECTED
│
├─ PSI or statistical test significant?
│  ├─ No → Return to regular monitoring
│  └─ Yes → Continue
│
├─ Can you measure ground truth on recent data?
│  ├─ Yes → Measure model accuracy on recent labeled data
│  │   ├─ Accuracy similar to baseline? → Covariate shift (monitor)
│  │   └─ Accuracy degraded? → Concept drift (retrain)
│  └─ No → Estimate based on feature drift
│     ├─ Feature drift is small (PSI <0.5) → Monitor
│     └─ Feature drift is large (PSI >0.5) → Retrain
│
├─ Are there known reasons for drift (new business rules, product changes)?
│  ├─ Yes → Understand the change, retrain if needed, document
│  └─ No → Investigate upstream systems
│
└─ DECISION: Monitor | Retrain | Investigate | Escalate

Case Study: Concept Drift in Customer Service AI

Scenario

System: AI chatbot for customer support. Trained on 2023 support tickets to classify intent (refund request, technical issue, billing question, feedback).

Model: Text classifier using TF-IDF + logistic regression. Baseline accuracy: 94%.

Monitoring setup: Daily performance tracking. Monthly PSI analysis on input features.

Timeline

June 2024 (Month 1-2): Model performs well. Accuracy 93-94%. No drift signals.

August 2024 (Month 3): Company launches new product. Customer queries shift. New intent type emerges: "feature request" for the new product. Model lumps these into "feedback" because it wasn't trained on feature request language.

Detection: PSI on text features shows increase to 0.18 (small drift). But more importantly, manual review reveals 15% of tickets are misclassified (misclassification rate 5% → 20%).

Investigation: Compare baseline (pre-launch) and recent (post-launch) tickets. Find 200 feature request tickets that model classified as feedback. These tickets use language (e.g., "can you add...", "would it be possible to...") that wasn't common in training data.

Root cause: Concept drift. The meaning of "feedback" has shifted. New intent category emerged.

Response: Collect 500 labeled tickets from August-September (including new feature requests). Add "feature request" as 5th intent class. Retrain model on 2023 data + new labeled data.

Result: Model accuracy on new data: 96%. Misclassification rate back to 4%.

Lesson: Concept drift often correlates with business changes. Monitor both data distributions and business context. When new product launches, retrain preemptively.

Summary and Monitoring Strategy