Bayesian Change-Point Detection
What you'll learn: How to use Bayesian inference to find where your data fundamentally changes behavior—and express your confidence about those locations.
The Core Idea
Bayesian Change-Point Detection treats the location of changes in your data as uncertain quantities rather than single fixed points. Instead of saying "the change happened exactly at time 42," it says "the change most likely happened around time 42, with this probability distribution."
Think of it like a detective investigating when someone left a party. Rather than pinpointing one exact moment, Bayesian reasoning considers all the clues (data) and says: "Based on the evidence, they probably left between 10:15 and 10:30 PM, with 10:22 PM being most likely."
How It Works
You start with a prior belief about where change-points might occur (uniform across time, or perhaps favoring certain periods). As you observe data, you update this belief using Bayes' theorem, combining:
- Likelihood: How well different change-point locations explain the observed data patterns
- Prior: Your initial assumptions about change-point locations
- Posterior: The updated probability distribution over possible change-point locations
The output isn't just "change at index 150"—it's a probability distribution showing the likelihood of changes at every position. You also get credible intervals (e.g., "95% confident the change occurred between indices 145-155").
Why It Matters
Unlike PELT or CUSUM, Bayesian methods naturally quantify uncertainty. This is crucial when you need to know not just where changes likely occurred, but how confident you should be. It handles noisy data gracefully and can incorporate domain knowledge through priors.
Key Takeaway: Bayesian Change-Point Detection provides a probabilistic framework that estimates both the locations of changes and the uncertainty around those estimates, making it ideal when confidence quantification matters as much as detection itself.