Struct BayesianChangePoint 

pub struct BayesianChangePoint {
    pub hazard_rate: f64,
    pub prior_mean: f64,
    pub prior_var: f64,
    pub obs_var: f64,
    pub threshold: f64,
}

Bayesian Online Change Point Detection.

Implements the Adams & MacKay (2007) algorithm that maintains a distribution over “run lengths” (time since last change point). At each time step:

1. Observe new data point x_t
2. Update run length probabilities using predictive likelihood
3. Apply hazard function to compute growth and change point probabilities

The hazard function H(tau) gives the probability of a change point given the current run length tau. A constant hazard 1/lambda yields a geometric prior on run lengths.

References

Adams, R. P., & MacKay, D. J. C. (2007). “Bayesian online changepoint detection.” arXiv preprint arXiv:0710.3742.

Fields

hazard_rate: f64

Hazard rate (1/lambda): probability of change at each time step

prior_mean: f64

Prior mean for the normal model

prior_var: f64

Prior variance for the normal model

obs_var: f64

Observation noise variance

threshold: f64

Threshold on change point posterior probability

Implementations

impl BayesianChangePoint

pub fn new(hazard_rate: f64, threshold: f64) -> Self

Create a new Bayesian change point detector.

Arguments

• hazard_rate - Constant hazard rate (1/lambda). Higher values mean change points are more frequent a priori.
• threshold - Posterior probability threshold for declaring a change point.

pub fn with_prior( hazard_rate: f64, threshold: f64, prior_mean: f64, prior_var: f64, obs_var: f64, ) -> Self

Create with custom prior parameters.

pub fn detect(&self, data: &ArrayView1<'_, f64>) -> Result<ChangePointResult>

Detect change points using Bayesian online detection.

Uses the MAP (Maximum A Posteriori) run length to detect changepoints. With a constant hazard function, P(r_t = 0 | x_{1:t}) = h always, so the raw run-length-zero probability cannot distinguish changepoints. Instead, we detect times when the MAP run length drops significantly, indicating the run length distribution has shifted to short run lengths after a structural break in the data.

The threshold parameter controls the minimum posterior probability of the new MAP run length required to declare a changepoint, filtering out low-confidence detections.

pub fn compute_run_length_distribution( &self, data: &ArrayView1<'_, f64>, ) -> Result<(Array2<f64>, Array1<f64>)>

Compute the full run length distribution over time.

Returns:

• run_length_probs - Matrix where row t contains the probability of each run length at time t. Entry [t][r] = P(r_t = r | x_{1:t}).
• cp_probs - Vector of change point posterior probabilities at each time step.

pub fn hazard(&self, _run_length: usize) -> f64

Constant hazard function value.

Trait Implementations

impl Clone for BayesianChangePoint

fn clone(&self) -> BayesianChangePoint

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for BayesianChangePoint

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.