Model Evaluation

name: model-evaluation description: Evaluates machine learning models for performance, fairness, and reliability using appropriate metrics and validation techniques. Covers training debugging, hyperparameter tuning, and production monitoring. Trigger keywords: model evaluation, metrics, accuracy, precision, recall, F1, F1-score, ROC, AUC, ROC-AUC, confusion matrix, cross-validation, k-fold, stratified, overfitting, underfitting, bias, variance, bias-variance tradeoff, hyperparameter, hyperparameter tuning, loss, loss function, metric, benchmark, benchmarking, model performance, classification metrics, regression metrics, RMSE, MSE, MAE, MAPE, R2, R-squared, train-test split, validation set, test set, hold-out, learning curve, validation curve, model selection, error analysis, residual analysis, ML testing, training issues, convergence, gradient, vanishing gradient, exploding gradient, training instability, LLM evaluation, language model evaluation, prompt engineering evaluation, A/B testing, champion-challenger, model monitoring, model drift, data drift, concept drift, model decay. allowed-tools: Read, Grep, Glob, Edit, Write, Bash

Model Evaluation

Overview

This skill focuses on comprehensive evaluation of machine learning models across the entire ML lifecycle. It covers metric selection, validation strategies, fairness assessment, training debugging, hyperparameter tuning, LLM evaluation, A/B testing, and production monitoring for ensuring model quality and reliability.

When to Use This Skill

Use this skill when you need to:

• Select and compute appropriate evaluation metrics for ML models
• Design cross-validation strategies and train/test splits
• Debug training issues (overfitting, underfitting, convergence problems)
• Tune hyperparameters and validate model improvements
• Evaluate LLMs and generative models
• Conduct A/B tests for model comparison in production
• Monitor deployed models for drift and degradation
• Assess model fairness across demographic groups
• Analyze error patterns and residuals
• Create evaluation reports and dashboards

Related Skills and Agents

When to Escalate:

• senior-software-engineer (Opus): For ML system architecture decisions, model selection strategies, complex evaluation pipeline design
• security-engineer (Opus): For adversarial robustness evaluation, model poisoning detection, security-aware metrics
• senior-infrastructure-engineer (Opus): For distributed evaluation infrastructure, large-scale benchmarking, production monitoring architecture

Complementary Skills:

• Use /debugging for systematic debugging of evaluation pipelines and metric computation issues
• Use /testing for unit testing evaluation code and validation logic
• Use /data-validation for input data quality checks before model evaluation

Instructions

1. Define Evaluation Criteria

Business Alignment:

• Identify business objectives and success criteria
• Translate business goals to ML metrics
• Define acceptable performance thresholds
• Consider cost of different error types (false positives vs false negatives)

Metric Selection:

• Classification: accuracy, precision, recall, F1, ROC-AUC, PR-AUC
• Regression: MSE, RMSE, MAE, MAPE, R2, explained variance
• Ranking: NDCG, MAP, MRR
• LLMs: perplexity, BLEU, ROUGE, BERTScore, human eval
• Custom metrics for domain-specific requirements

Fairness Requirements:

• Identify protected attributes (race, gender, age)
• Choose fairness definitions (demographic parity, equalized odds)
• Set fairness constraints and thresholds

2. Design Evaluation Strategy

Data Splitting:

• Train/validation/test split ratios (e.g., 60/20/20)
• Stratified splits for class imbalance
• Time-based splits for temporal data
• Group-based splits to prevent data leakage

Cross-Validation:

• K-fold CV for standard problems
• Stratified K-fold for imbalanced classes
• TimeSeriesSplit for temporal data
• GroupKFold for clustered data
• Leave-one-out for small datasets

Handling Imbalance:

• Stratified sampling
• Class weights in metrics
• Resampling strategies (SMOTE, undersampling)
• Appropriate metrics (F1, PR-AUC instead of accuracy)

3. Conduct Evaluation

Performance Metrics:

• Calculate primary and secondary metrics
• Compute confidence intervals
• Compare against baselines
• Statistical significance testing

Error Analysis:

• Confusion matrix analysis
• Per-class performance breakdown
• Error type categorization
• Hard example mining

Fairness Assessment:

• Group-wise metric comparison
• Demographic parity evaluation
• Equalized odds analysis
• Disparate impact measurement

Edge Case Testing:

• Boundary condition validation
• Out-of-distribution detection
• Adversarial robustness
• Stress testing with extreme inputs

4. Debug Training Issues

Overfitting Detection:

• Train vs validation performance gap
• Learning curves analysis
• Validation metrics plateauing while training improves
• Mitigation: regularization, dropout, early stopping, data augmentation

Underfitting Detection:

• Poor performance on both train and validation
• Learning curves not converging
• Model too simple for problem complexity
• Mitigation: increase model capacity, feature engineering, longer training

Convergence Problems:

• Loss not decreasing
• Loss oscillating or unstable
• Exploding gradients (loss becomes NaN)
• Vanishing gradients (loss stays constant)
• Mitigation: learning rate tuning, gradient clipping, batch normalization

Learning Rate Issues:

• Too high: training unstable, loss oscillates
• Too low: training too slow, stuck in local minima
• Solution: learning rate schedules, warmup, cosine annealing

Batch Size Effects:

• Small batch: noisy gradients, poor generalization
• Large batch: memory issues, sharp minima
• Find sweet spot through experimentation

5. Hyperparameter Tuning

Search Strategies:

• Grid search: exhaustive but expensive
• Random search: better coverage for high-dimensional spaces
• Bayesian optimization: sample efficient
• Hyperband: adaptive resource allocation

Key Hyperparameters:

• Learning rate (most critical)
• Batch size
• Regularization strength (L1, L2, dropout)
• Network architecture (layers, units)
• Optimizer choice (Adam, SGD, AdamW)

Validation:

• Use validation set for hyperparameter selection
• Never tune on test set
• Consider nested cross-validation for small datasets

6. LLM and Generative Model Evaluation

Automatic Metrics:

• Perplexity for language models
• BLEU, ROUGE for text generation
• BERTScore for semantic similarity
• Exact match, F1 for QA tasks

Human Evaluation:

• Fluency, coherence, relevance
• Factual accuracy
• Safety and toxicity
• Instruction following

Prompt Engineering Evaluation:

• Few-shot vs zero-shot comparison
• Prompt template A/B testing
• Chain-of-thought effectiveness
• System message impact

LLM-as-Judge:

• Use stronger models to evaluate weaker models
• Pairwise comparison for ranking
• Rubric-based scoring
• Calibration against human judgments

7. A/B Testing for Model Comparison

Experimental Design:

• Random traffic split (50/50 or 90/10)
• Minimum sample size calculation
• Statistical power analysis
• Duration planning for seasonality

Metrics:

• Primary business metric (conversion, revenue)
• Secondary metrics (latency, user satisfaction)
• Guardrail metrics (error rate, bias)
• Sample ratio mismatch checks

Analysis:

• Statistical significance testing (t-test, Mann-Whitney)
• Effect size estimation
• Confidence intervals
• Multiple testing correction (Bonferroni)

Decision Criteria:

• Primary metric improvement threshold
• No degradation in guardrail metrics
• Sufficient statistical power
• Business case validation

8. Production Model Monitoring

Performance Monitoring:

• Track key metrics over time
• Compare against baseline/champion model
• Detect performance degradation
• Alert on threshold violations

Data Drift Detection:

• Input distribution shifts
• Feature statistics tracking
• KL divergence, KS test, PSI
• Covariate shift detection

Concept Drift Detection:

• Model prediction distribution changes
• Label distribution shifts (when available)
• Performance metric trends
• Adversarial Validation

Monitoring Infrastructure:

• Real-time metric computation
• Dashboards for visualization
• Alerting and on-call rotation
• Automated retraining triggers

9. Report and Document

Evaluation Report Structure:

• Executive summary with key findings
• Methodology and experimental setup
• Comprehensive metric tables
• Error analysis and case studies
• Fairness assessment results
• Recommendations and next steps

Visualization:

• ROC and PR curves
• Confusion matrices
• Learning curves
• Residual plots
• Fairness comparison charts

Version Control:

• Model version and checkpoints
• Dataset versions and splits
• Hyperparameter configurations
• Evaluation code and environment

Best Practices

General Principles

1. Match Metrics to Goals: Choose metrics aligned with business objectives, not just academic standards
2. Use Multiple Metrics: No single metric tells the whole story; use complementary metrics
3. Proper Validation: Use appropriate cross-validation schemes to avoid overfitting to validation set
4. Test Distribution Shift: Evaluate on out-of-distribution data to assess generalization
5. Check for Bias: Assess fairness across demographic groups before deployment
6. Version Everything: Track models, data, metrics, and code for reproducibility
7. Monitor Production: Continuously track model performance after deployment

Training and Debugging

1. Start Simple: Begin with simple baselines before complex models
2. Visualize Learning: Plot learning curves early and often
3. Debug Incrementally: Change one thing at a time when debugging training issues
4. Sanity Check: Overfit on small batch first to verify model can learn
5. Early Stopping: Use validation-based early stopping to prevent overfitting
6. Gradient Monitoring: Track gradient norms to detect vanishing/exploding gradients

Evaluation Rigor

1. Hold-out Test Set: Never touch test set until final evaluation
2. Stratified Splits: Use stratification for imbalanced datasets
3. Statistical Testing: Use significance tests for model comparisons
4. Error Analysis: Dive deep into errors to understand failure modes
5. Temporal Validation: For time-series, validate on future data only

Production and Monitoring

1. Shadow Mode: Deploy new models in shadow mode before switching traffic
2. Gradual Rollout: Use canary deployments or gradual percentage rollouts
3. Rollback Plan: Have automated rollback triggers for performance degradation
4. Alert Fatigue: Set meaningful alert thresholds to avoid noise

Examples

Example 1: Classification Model Evaluation

import numpy as np
import pandas as pd
from sklearn.metrics import (
    accuracy_score, precision_score, recall_score, f1_score,
    roc_auc_score, average_precision_score, confusion_matrix,
    classification_report, roc_curve, precision_recall_curve
)
import matplotlib.pyplot as plt

class ClassificationEvaluator:
    """Comprehensive classification model evaluator."""

    def __init__(self, y_true, y_pred, y_prob=None, class_names=None):
        self.y_true = y_true
        self.y_pred = y_pred
        self.y_prob = y_prob
        self.class_names = class_names or ['Negative', 'Positive']

    def compute_metrics(self) -> dict:
        """Compute all classification metrics."""
        metrics = {
            'accuracy': accuracy_score(self.y_true, self.y_pred),
            'precision': precision_score(self.y_true, self.y_pred, average='weighted'),
            'recall': recall_score(self.y_true, self.y_pred, average='weighted'),
            'f1': f1_score(self.y_true, self.y_pred, average='weighted'),
        }

        if self.y_prob is not None:
            metrics['roc_auc'] = roc_auc_score(self.y_true, self.y_prob)
            metrics['average_precision'] = average_precision_score(self.y_true, self.y_prob)

        return metrics

    def confusion_matrix_analysis(self) -> dict:
        """Analyze confusion matrix in detail."""
        cm = confusion_matrix(self.y_true, self.y_pred)
        tn, fp, fn, tp = cm.ravel()

        return {
            'confusion_matrix': cm,
            'true_negatives': tn,
            'false_positives': fp,
            'false_negatives': fn,
            'true_positives': tp,
            'specificity': tn / (tn + fp),
            'sensitivity': tp / (tp + fn),
            'false_positive_rate': fp / (fp + tn),
            'false_negative_rate': fn / (fn + tp),
        }

    def plot_roc_curve(self, save_path=None):
        """Plot ROC curve with AUC."""
        if self.y_prob is None:
            raise ValueError("Probabilities required for ROC curve")

        fpr, tpr, thresholds = roc_curve(self.y_true, self.y_prob)
        auc = roc_auc_score(self.y_true, self.y_prob)

        plt.figure(figsize=(8, 6))
        plt.plot(fpr, tpr, label=f'ROC (AUC = {auc:.3f})')
        plt.plot([0, 1], [0, 1], 'k--', label='Random')
        plt.xlabel('False Positive Rate')
        plt.ylabel('True Positive Rate')
        plt.title('Receiver Operating Characteristic (ROC) Curve')
        plt.legend()
        plt.grid(True, alpha=0.3)

        if save_path:
            plt.savefig(save_path, dpi=150, bbox_inches='tight')
        plt.show()

    def generate_report(self) -> str:
        """Generate comprehensive evaluation report."""
        metrics = self.compute_metrics()
        cm_analysis = self.confusion_matrix_analysis()

        report = f"""
# Classification Model Evaluation Report

## Overall Metrics
| Metric | Value |
|--------|-------|
| Accuracy | {metrics['accuracy']:.4f} |
| Precision | {metrics['precision']:.4f} |
| Recall | {metrics['recall']:.4f} |
| F1 Score | {metrics['f1']:.4f} |
| ROC AUC | {metrics.get('roc_auc', 'N/A'):.4f if isinstance(metrics.get('roc_auc'), float) else 'N/A'} |

## Confusion Matrix Analysis
| Metric | Value |
|--------|-------|
| True Positives | {cm_analysis['true_positives']} |
| True Negatives | {cm_analysis['true_negatives']} |
| False Positives | {cm_analysis['false_positives']} |
| False Negatives | {cm_analysis['false_negatives']} |
| Sensitivity | {cm_analysis['sensitivity']:.4f} |
| Specificity | {cm_analysis['specificity']:.4f} |

## Detailed Classification Report
{classification_report(self.y_true, self.y_pred, target_names=self.class_names)}
"""
        return report

# Usage
evaluator = ClassificationEvaluator(y_true, y_pred, y_prob)
print(evaluator.generate_report())
evaluator.plot_roc_curve('roc_curve.png')

Example 2: Regression Model Evaluation

from sklearn.metrics import (
    mean_squared_error, mean_absolute_error, r2_score,
    mean_absolute_percentage_error, explained_variance_score
)
import numpy as np

class RegressionEvaluator:
    """Comprehensive regression model evaluator."""

    def __init__(self, y_true, y_pred):
        self.y_true = np.array(y_true)
        self.y_pred = np.array(y_pred)
        self.residuals = self.y_true - self.y_pred

    def compute_metrics(self) -> dict:
        """Compute all regression metrics."""
        mse = mean_squared_error(self.y_true, self.y_pred)

        return {
            'mse': mse,
            'rmse': np.sqrt(mse),
            'mae': mean_absolute_error(self.y_true, self.y_pred),
            'mape': mean_absolute_percentage_error(self.y_true, self.y_pred) * 100,
            'r2': r2_score(self.y_true, self.y_pred),
            'explained_variance': explained_variance_score(self.y_true, self.y_pred),
        }

    def residual_analysis(self) -> dict:
        """Analyze residual patterns."""
        return {
            'mean_residual': np.mean(self.residuals),
            'std_residual': np.std(self.residuals),
            'max_overestimate': np.min(self.residuals),
            'max_underestimate': np.max(self.residuals),
            'residual_skewness': self._skewness(self.residuals),
        }

    def _skewness(self, data):
        """Calculate skewness."""
        n = len(data)
        mean = np.mean(data)
        std = np.std(data)
        return (n / ((n-1) * (n-2))) * np.sum(((data - mean) / std) ** 3)

    def plot_diagnostics(self, save_path=None):
        """Plot diagnostic plots for residual analysis."""
        fig, axes = plt.subplots(2, 2, figsize=(12, 10))

        # Actual vs Predicted
        ax1 = axes[0, 0]
        ax1.scatter(self.y_true, self.y_pred, alpha=0.5)
        ax1.plot([self.y_true.min(), self.y_true.max()],
                 [self.y_true.min(), self.y_true.max()], 'r--')
        ax1.set_xlabel('Actual')
        ax1.set_ylabel('Predicted')
        ax1.set_title('Actual vs Predicted')

        # Residuals vs Predicted
        ax2 = axes[0, 1]
        ax2.scatter(self.y_pred, self.residuals, alpha=0.5)
        ax2.axhline(y=0, color='r', linestyle='--')
        ax2.set_xlabel('Predicted')
        ax2.set_ylabel('Residuals')
        ax2.set_title('Residuals vs Predicted')

        # Residual histogram
        ax3 = axes[1, 0]
        ax3.hist(self.residuals, bins=30, edgecolor='black')
        ax3.set_xlabel('Residual')
        ax3.set_ylabel('Frequency')
        ax3.set_title('Residual Distribution')

        # Q-Q plot
        ax4 = axes[1, 1]
        from scipy import stats
        stats.probplot(self.residuals, dist="norm", plot=ax4)
        ax4.set_title('Q-Q Plot')

        plt.tight_layout()
        if save_path:
            plt.savefig(save_path, dpi=150, bbox_inches='tight')
        plt.show()

Example 3: Cross-Validation Strategies

from sklearn.model_selection import (
    cross_val_score, StratifiedKFold, TimeSeriesSplit,
    GroupKFold, cross_validate
)

def evaluate_with_cv(model, X, y, cv_strategy='stratified', n_splits=5, groups=None):
    """
    Evaluate model with appropriate cross-validation strategy.

    Args:
        model: Sklearn-compatible model
        X: Features
        y: Target
        cv_strategy: 'stratified', 'timeseries', 'group', or 'kfold'
        n_splits: Number of CV folds
        groups: Group labels for GroupKFold

    Returns:
        Dictionary with CV results
    """
    # Select CV strategy
    if cv_strategy == 'stratified':
        cv = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=42)
    elif cv_strategy == 'timeseries':
        cv = TimeSeriesSplit(n_splits=n_splits)
    elif cv_strategy == 'group':
        cv = GroupKFold(n_splits=n_splits)
    else:
        cv = n_splits

    # Define scoring metrics
    scoring = {
        'accuracy': 'accuracy',
        'precision': 'precision_weighted',
        'recall': 'recall_weighted',
        'f1': 'f1_weighted',
        'roc_auc': 'roc_auc'
    }

    # Perform cross-validation
    cv_results = cross_validate(
        model, X, y,
        cv=cv,
        scoring=scoring,
        groups=groups,
        return_train_score=True,
        n_jobs=-1
    )

    # Summarize results
    summary = {}
    for metric in scoring.keys():
        test_scores = cv_results[f'test_{metric}']
        train_scores = cv_results[f'train_{metric}']
        summary[metric] = {
            'test_mean': np.mean(test_scores),
            'test_std': np.std(test_scores),
            'train_mean': np.mean(train_scores),
            'train_std': np.std(train_scores),
            'overfit_gap': np.mean(train_scores) - np.mean(test_scores)
        }

    return summary

# Usage example
results = evaluate_with_cv(model, X, y, cv_strategy='stratified', n_splits=5)
for metric, values in results.items():
    print(f"{metric}: {values['test_mean']:.4f} (+/- {values['test_std']:.4f})")
    print(f"  Overfitting gap: {values['overfit_gap']:.4f}")

Example 4: Fairness Evaluation

def evaluate_fairness(y_true, y_pred, sensitive_attr, favorable_label=1):
    """
    Evaluate model fairness across demographic groups.

    Args:
        y_true: True labels
        y_pred: Predicted labels
        sensitive_attr: Protected attribute values
        favorable_label: The favorable outcome label

    Returns:
        Dictionary with fairness metrics
    """
    groups = np.unique(sensitive_attr)
    results = {'group_metrics': {}}

    for group in groups:
        mask = sensitive_attr == group
        group_true = y_true[mask]
        group_pred = y_pred[mask]

        # Calculate group-specific metrics
        tp = np.sum((group_true == favorable_label) & (group_pred == favorable_label))
        fp = np.sum((group_true != favorable_label) & (group_pred == favorable_label))
        fn = np.sum((group_true == favorable_label) & (group_pred != favorable_label))
        tn = np.sum((group_true != favorable_label) & (group_pred != favorable_label))

        results['group_metrics'][group] = {
            'selection_rate': np.mean(group_pred == favorable_label),
            'tpr': tp / (tp + fn) if (tp + fn) > 0 else 0,
            'fpr': fp / (fp + tn) if (fp + tn) > 0 else 0,
            'accuracy': np.mean(group_true == group_pred),
            'size': len(group_true)
        }

    # Calculate fairness metrics
    selection_rates = [m['selection_rate'] for m in results['group_metrics'].values()]
    tprs = [m['tpr'] for m in results['group_metrics'].values()]
    fprs = [m['fpr'] for m in results['group_metrics'].values()]

    results['fairness_metrics'] = {
        'demographic_parity_diff': max(selection_rates) - min(selection_rates),
        'equalized_odds_tpr_diff': max(tprs) - min(tprs),
        'equalized_odds_fpr_diff': max(fprs) - min(fprs),
    }

    return results

Example 5: Training Debugging with Learning Curves

import matplotlib.pyplot as plt
from sklearn.model_selection import learning_curve

def plot_learning_curves(model, X, y, cv=5, train_sizes=np.linspace(0.1, 1.0, 10)):
    """
    Plot learning curves to diagnose overfitting/underfitting.

    Args:
        model: Sklearn-compatible model
        X: Features
        y: Target
        cv: Cross-validation folds
        train_sizes: Array of training set size fractions
    """
    train_sizes, train_scores, val_scores = learning_curve(
        model, X, y,
        cv=cv,
        train_sizes=train_sizes,
        scoring='accuracy',
        n_jobs=-1
    )

    train_mean = np.mean(train_scores, axis=1)
    train_std = np.std(train_scores, axis=1)
    val_mean = np.mean(val_scores, axis=1)
    val_std = np.std(val_scores, axis=1)

    plt.figure(figsize=(10, 6))
    plt.plot(train_sizes, train_mean, label='Training score', color='blue', marker='o')
    plt.fill_between(train_sizes, train_mean - train_std, train_mean + train_std, alpha=0.15, color='blue')
    plt.plot(train_sizes, val_mean, label='Validation score', color='red', marker='o')
    plt.fill_between(train_sizes, val_mean - val_std, val_mean + val_std, alpha=0.15, color='red')

    plt.xlabel('Training Set Size')
    plt.ylabel('Accuracy')
    plt.title('Learning Curves')
    plt.legend(loc='lower right')
    plt.grid(True, alpha=0.3)

    # Add diagnostic annotations
    final_gap = train_mean[-1] - val_mean[-1]
    if final_gap > 0.1:
        plt.text(0.5, 0.05, 'HIGH OVERFITTING: Large gap between train and validation',
                 transform=plt.gca().transAxes, color='red', fontweight='bold')
    elif val_mean[-1] < 0.7:
        plt.text(0.5, 0.05, 'UNDERFITTING: Both train and validation scores are low',
                 transform=plt.gca().transAxes, color='orange', fontweight='bold')

    plt.tight_layout()
    plt.show()

    return {
        'final_train_score': train_mean[-1],
        'final_val_score': val_mean[-1],
        'overfit_gap': final_gap
    }

Example 6: LLM Evaluation

from typing import List, Dict
import openai

def evaluate_llm_generation(
    prompts: List[str],
    references: List[str],
    model: str,
    judge_model: str = "gpt-4"
) -> Dict:
    """
    Evaluate LLM generation quality using LLM-as-judge.

    Args:
        prompts: Input prompts
        references: Reference outputs (if available)
        model: Model to evaluate
        judge_model: Model to use as judge

    Returns:
        Dictionary with evaluation scores
    """
    results = []

    for prompt, reference in zip(prompts, references):
        # Generate response
        response = openai.ChatCompletion.create(
            model=model,
            messages=[{"role": "user", "content": prompt}]
        )
        generation = response.choices[0].message.content

        # LLM-as-judge evaluation
        judge_prompt = f"""Evaluate the following AI-generated response on a scale of 1-5 for:
1. Accuracy: Is the information correct?
2. Relevance: Does it address the prompt?
3. Fluency: Is it well-written and coherent?
4. Helpfulness: Is it useful to the user?

Prompt: {prompt}
Reference (if available): {reference}
Response: {generation}

Provide scores in JSON format: {{"accuracy": X, "relevance": X, "fluency": X, "helpfulness": X, "overall": X}}
"""

        judge_response = openai.ChatCompletion.create(
            model=judge_model,
            messages=[{"role": "user", "content": judge_prompt}]
        )

        scores = json.loads(judge_response.choices[0].message.content)
        results.append({
            'prompt': prompt,
            'generation': generation,
            'scores': scores
        })

    # Aggregate scores
    avg_scores = {}
    for key in ['accuracy', 'relevance', 'fluency', 'helpfulness', 'overall']:
        avg_scores[key] = np.mean([r['scores'][key] for r in results])

    return {
        'individual_results': results,
        'average_scores': avg_scores
    }

Example 7: A/B Test Analysis

from scipy import stats

def analyze_ab_test(control_metric: np.ndarray, treatment_metric: np.ndarray, alpha: float = 0.05):
    """
    Analyze A/B test results with statistical significance testing.

    Args:
        control_metric: Metric values for control group
        treatment_metric: Metric values for treatment group
        alpha: Significance level

    Returns:
        Dictionary with test results
    """
    # Descriptive statistics
    control_mean = np.mean(control_metric)
    treatment_mean = np.mean(treatment_metric)
    relative_lift = (treatment_mean - control_mean) / control_mean * 100

    # Statistical test
    t_stat, p_value = stats.ttest_ind(treatment_metric, control_metric)
    is_significant = p_value < alpha

    # Effect size (Cohen's d)
    pooled_std = np.sqrt((np.std(control_metric)**2 + np.std(treatment_metric)**2) / 2)
    cohens_d = (treatment_mean - control_mean) / pooled_std

    # Confidence interval
    ci = stats.t.interval(
        confidence=1-alpha,
        df=len(control_metric) + len(treatment_metric) - 2,
        loc=treatment_mean - control_mean,
        scale=stats.sem(np.concatenate([control_metric, treatment_metric]))
    )

    return {
        'control_mean': control_mean,
        'treatment_mean': treatment_mean,
        'relative_lift_pct': relative_lift,
        'p_value': p_value,
        'is_significant': is_significant,
        'cohens_d': cohens_d,
        'confidence_interval': ci,
        'recommendation': 'LAUNCH' if is_significant and relative_lift > 0 else 'DO NOT LAUNCH'
    }

# Usage
results = analyze_ab_test(control_conversions, treatment_conversions)
print(f"Relative Lift: {results['relative_lift_pct']:.2f}%")
print(f"P-value: {results['p_value']:.4f}")
print(f"Recommendation: {results['recommendation']}")

Example 8: Production Model Monitoring

from scipy.stats import ks_2samp
import pandas as pd

class ModelMonitor:
    """Monitor deployed model for drift and degradation."""

    def __init__(self, baseline_data: pd.DataFrame, baseline_predictions: np.ndarray):
        self.baseline_data = baseline_data
        self.baseline_predictions = baseline_predictions

    def detect_data_drift(self, current_data: pd.DataFrame, threshold: float = 0.05) -> Dict:
        """Detect feature distribution drift using KS test."""
        drift_results = {}

        for col in self.baseline_data.columns:
            if pd.api.types.is_numeric_dtype(self.baseline_data[col]):
                statistic, p_value = ks_2samp(
                    self.baseline_data[col].dropna(),
                    current_data[col].dropna()
                )
                drift_results[col] = {
                    'ks_statistic': statistic,
                    'p_value': p_value,
                    'drift_detected': p_value < threshold
                }

        return drift_results

    def detect_prediction_drift(self, current_predictions: np.ndarray, threshold: float = 0.05) -> Dict:
        """Detect prediction distribution drift."""
        statistic, p_value = ks_2samp(self.baseline_predictions, current_predictions)

        return {
            'ks_statistic': statistic,
            'p_value': p_value,
            'drift_detected': p_value < threshold,
            'baseline_mean': np.mean(self.baseline_predictions),
            'current_mean': np.mean(current_predictions),
            'mean_shift': np.mean(current_predictions) - np.mean(self.baseline_predictions)
        }

    def performance_degradation_check(
        self,
        current_metric: float,
        baseline_metric: float,
        threshold_pct: float = 5.0
    ) -> Dict:
        """Check for performance degradation."""
        degradation_pct = (baseline_metric - current_metric) / baseline_metric * 100

        return {
            'baseline_metric': baseline_metric,
            'current_metric': current_metric,
            'degradation_pct': degradation_pct,
            'alert': degradation_pct > threshold_pct,
            'recommendation': 'RETRAIN MODEL' if degradation_pct > threshold_pct else 'OK'
        }

# Usage
monitor = ModelMonitor(baseline_df, baseline_preds)
drift_check = monitor.detect_data_drift(current_df)
pred_drift = monitor.detect_prediction_drift(current_preds)
perf_check = monitor.performance_degradation_check(current_accuracy, baseline_accuracy)

Common Pitfalls

1. Test Set Contamination: Never use test set for hyperparameter tuning or model selection
2. Data Leakage: Ensure validation/test data doesn't leak into training (temporal ordering, group splits)
3. Wrong Metric Choice: Using accuracy for imbalanced datasets, not considering business costs
4. Ignoring Confidence Intervals: Point estimates without uncertainty can be misleading
5. Multiple Comparisons: Not correcting p-values when testing many hypotheses
6. Survivorship Bias: Evaluating only on successful cases, ignoring failures
7. Overfitting to Validation: Repeatedly tuning on validation set effectively makes it a second training set
8. Ignoring Fairness: Deploying models without fairness evaluation can cause harm
9. No Baseline: Not comparing against simple baselines (random, majority class, linear model)
10. Production-Training Skew: Evaluation setup doesn't match production environment

Additional Resources

• Metrics: Scikit-learn metrics documentation, Hugging Face evaluate library
• Fairness: AI Fairness 360, Fairlearn
• LLM Evaluation: HELM, lm-evaluation-harness, BIG-bench
• A/B Testing: Evan Miller's A/B testing tools, experimentation platform docs
• Monitoring: Evidently AI, WhyLabs, Fiddler