Skip to content

Statistical Models

Pure statistical approaches without machine learning.

Overview

Model ROC AUC F1 Score Train Time
Hypothesis Testing 0.5394 0.4167 0s
Bayesian BOCPD 0.5005 0.0625 183s
welch_ttest 0.4634 0.0000 0s

Key Advantage: No Training

Training-Free

These models don't require training data. They use statistical theory to detect breaks directly.

Use Cases

When to Use Statistical Models

  • Baseline comparison — Benchmark for ML models
  • Interpretability required — p-values and test statistics are understandable
  • No training data — Deploy immediately without labeled examples
  • Real-time streaming — Low latency, no model loading

When to Avoid

  • Higher AUC needed — ML models achieved higher robust scores
  • Complex patterns — Statistical tests assume specific distributions

Theory Behind the Tests

Hypothesis Testing Framework

\[ H_0: \text{No structural break (same distribution)} $$ $$ H_1: \text{Structural break exists (different distributions)} \]

Key Test Statistics

Test What It Measures
Welch's t-test Mean difference significance
Kolmogorov-Smirnov Maximum CDF difference
Mann-Whitney U Rank-based comparison
Levene's test Variance equality
CUSUM Cumulative deviation from mean

Key Findings

Statistical Models as Baseline

hypothesis_testing_pure serves as a baseline showing the gap between pure statistics and ML approaches.

Tree-based models achieved higher AUC (0.7423 vs 0.5394) and robust scores than statistical methods.