When to Use BOCPD

Bayesian Online Changepoint Detection is most effective when you need probabilistic, real-time detection of regime shifts with calibrated uncertainty. This section contrasts BOCPD with other popular approaches.

Suitable Scenarios

Streaming/online monitoring – data arrive sequentially and decisions must be made on the fly (manufacturing sensors, telemetry, trading).
Piecewise-stationary processes – each regime can be modelled by a conjugate likelihood–prior pair from Conjugate Priors.
Low-latency alerts – changepoint probabilities \(P(r_t=0 \mid x_{1:t})\) provide interpretable confidence scores.
Limited data per segment – the Bayesian update works even with very few samples because the prior supplies regularisation.

Less Suitable Scenarios

Gradual drift – if changes are smooth rather than abrupt, methods such as Kalman filters or Prophet’s trend components may be preferable. There are extensions to BOCPD that handle gradual changes, but they are not implemented in Fast-BOCPD yet (e.g. different prior/hazard models).
Very high-dimensional observations – the current observation models are univariate; multivariate extensions demand custom conjugate pairs.

Comparison to Other Methods

Method	Strengths	Limitations vs BOCPD
CUSUM / Page-Hinkley	Simple, few parameters, O(1) memory	Tailored to mean shifts, requires handcrafted thresholds, no posterior
Prophet / ARIMA	Captures seasonality, trend, regressors	Not changepoint detectors per se; changepoints are model components, not probabilistic events
ruptures	Rich offline algorithms, many cost functions, visual tools	Batch only; recomputes from scratch when new data arrive
promised-ai / dtolpin BOCD	Bayesian online detection in other languages	Slower throughput (see Competitor Comparison)
HMMs / Switching Kalman filters	Explicit latent-state models, handle dependence	Require predefined number of states and transition matrices

Implementation Mapping

Observation models correspond to the likelihood/prior choices in the conjugate derivation. Users pick a class such as GaussianNIG or PoissonGamma rather than specifying raw pdfs.
Hazard functions encapsulate the gap distribution; the current ConstantHazard class matches the geometric prior used in the theory.
Outputs – BOCPD.update returns both the full posterior \(P(r_t \mid x_{1:t})\) and \(P(r_t=0 \mid x_{1:t})\). These probabilities can drive downstream alerting or feed into tools like OnlineChangeDetector (see API Reference).

If your application matches the assumptions above and you need calibrated, online decisions, Fast-BOCPD provides an efficient, theoretically grounded solution. For purely offline retrospective analysis or models with strong seasonality, use BOCPD in combination with other techniques (e.g. detrend with ARIMA, then apply BOCPD to residuals).