Reader question: When a financial or banking time series may change regime at an unknown time, how can an algorithm keep updating the probability that a break has just occurred while also representing uncertainty about how long the current regime has lasted?
Bayesian Online Changepoint Detection, usually abbreviated BOCPD, maintains a posterior distribution over run length: the number of observations since the most recent changepoint. Every new observation is evaluated under all currently plausible run lengths. Posterior mass either grows forward into a longer regime or resets toward run length zero when the new observation looks more compatible with a fresh segment.
This article owns the online run-length inference problem: ordered observations + segment model + prior parameters + changepoint hazard → run-length posterior, changepoint probability, predictive distribution, computational diagnostics and evidence about whether an apparent break is robust or merely an outlier/model mismatch.
It does not own hidden-state regime classification, CUSUM threshold design, general model-drift governance or automatic market decisions. A BOCPD posterior tells us how the data challenge a segmented generative model; it does not tell us why the change occurred.
This is public mathematical and computational education, not trading, investment or personalized financial advice.
1. The state variable is “time since the last change”
Let rt denote the run length at time t. If no changepoint occurs at t, then:
rt = rt−1 + 1.
If a changepoint occurs, then:
rt = 0.
Instead of choosing one run length immediately, BOCPD maintains:
P(rt | x1:t).
This distribution is the core output. It can be sharply concentrated on one long run, split across several possible changepoint locations, or suddenly move substantial mass toward zero.
2. Why the full posterior is more informative than one alarm bit
A binary detector says “change” or “no change.” BOCPD can say:
- there is 70% posterior mass on very recent run lengths;
- there are two plausible change locations;
- the most probable run length is 42 observations but uncertainty remains broad;
- the latest point was surprising, yet most mass still supports the old regime.
This richer state is useful for diagnosis. It also makes model assumptions visible because the posterior reacts through the predictive likelihood and hazard function.
3. The predictive likelihood
For every candidate previous run length r, the algorithm computes:
πt(r) = p(xt | rt−1=r, x1:t−1).
This asks: if the current segment really began r observations ago, how plausible is the new observation under the segment model fitted to that run?
With conjugate models, the predictive distribution can often be updated analytically using sufficient statistics for each run-length hypothesis. A Gaussian segment with uncertain mean and variance, for example, can be paired with conjugate priors that yield a Student-t posterior predictive distribution.
4. The hazard function supplies the prior chance of a break
Define H(r) as the prior probability that a changepoint occurs after a run of length r.
Then the transition from a previous run length r has two canonical branches:
P(rt=0 | rt−1=r) = H(r)
P(rt=r+1 | rt−1=r) = 1−H(r).
A constant hazard H(r)=h corresponds to a geometric prior over segment lengths. Its prior mean segment length is approximately 1/h under that convention.
The hazard is not learned merely by calling it “Bayesian.” It is a modelling choice and therefore a major sensitivity parameter.
5. Growth probability
For an existing run length r, the unnormalised probability of growing the segment is:
P(rt=r+1, x1:t) = P(rt−1=r, x1:t−1) × πt(r) × [1−H(r)].
If the new observation is well predicted by a long-running segment, the corresponding branch can preserve substantial posterior mass.
6. Changepoint probability
The reset branch aggregates over all previous run lengths:
P(rt=0, x1:t) = Σr P(rt−1=r, x1:t−1) × πt(r) × H(r).
After computing all growth and reset branches, normalise them to obtain the posterior over rt.
The algorithm is therefore a message-passing recursion: every plausible history sends probability mass either forward or back to zero.
7. A conceptual example
Imagine monitoring standardised residuals from a stable forecasting model. For many observations, the residuals fluctuate around zero with roughly constant scale. Posterior mass gradually moves toward long run lengths.
Now several consecutive observations arrive around +2 rather than 0. Under the long-run segment model, their predictive likelihood may be poor. Under a fresh segment, they may be more plausible. Posterior mass shifts toward short run lengths.
If the next observations return immediately to zero, the algorithm may reinterpret the episode as noise or an outlier. If the shifted mean persists, the short-run hypotheses gain continuing predictive support.
The mechanism is not “large observation = changepoint.” It is relative predictive evidence across competing run-length histories.
8. Inputs
Typical inputs include:
- ordered observations or model residuals;
- within-segment likelihood model;
- prior parameters for each new segment;
- hazard function H(r);
- initial run-length distribution;
- numerical representation, usually log space for stability;
- pruning or truncation policy;
- missing-data treatment;
- posterior summary or alert threshold if operational summaries are needed;
- versioned definitions of the monitored quantity.
9. Outputs
Useful outputs include:
- full posterior P(rt|x1:t);
- posterior probability of a very recent changepoint;
- MAP run length;
- posterior mean or quantiles of run length;
- one-step predictive distribution;
- posterior entropy or dispersion;
- probability mass removed by pruning or truncation;
- detected change candidates and their uncertainty;
- segment-parameter posterior summaries;
- sensitivity results across hazards and observation models.
10. Assumption: piecewise generative structure
Standard BOCPD is designed around segments within which the chosen generative parameters are stable enough for the segment model to make sense. Changepoints represent abrupt changes between those segments.
If the true process drifts smoothly every day, a sequence of abrupt-segment resets may be a poor representation.
Weak link: a model can detect the changepoints implied by its own segmentation assumptions even when reality is gradual rather than piecewise stationary.
11. Hazard misspecification changes posterior behaviour
A high constant hazard expects short regimes and gives reset branches more prior probability. A low hazard expects long regimes and resists breaks.
Falsifier: rerun the detector across a justified range of prior mean segment lengths. If the main changepoint disappears under modest hazard changes, the evidence is prior-sensitive and should be reported that way.
12. Outliers can masquerade as regime changes
A single extreme observation may be extremely unlikely under every long-run hypothesis. A fresh segment, whose prior predictive is often wider, can then receive a large relative advantage.
This can produce a temporary spike in changepoint probability even when the underlying regime did not change.
Robust observation models, heavy-tailed predictive distributions, explicit outlier states or confirmation through subsequent observations are possible responses.
Falsifier: inject isolated shocks into simulated no-change data. If one outlier routinely causes persistent regime resets, the observation model is too brittle for that noise process.
13. Heavy tails and volatility clustering matter
Financial residuals are often heavier-tailed and more volatility-clustered than a simple i.i.d. Gaussian segment model assumes.
If volatility rises but the mean does not change, a constant-variance BOCPD may see many observations as unexpectedly unlikely and infer repeated changepoints.
Falsifier: compare results under Gaussian versus heavy-tailed or volatility-aware segment models. If “regime breaks” vanish when realistic tails are introduced, the earlier detector was identifying distributional misspecification.
14. Serial dependence can create false persistence
Within-segment independence is another common simplifying assumption. Autocorrelated observations can generate runs of same-direction residuals that look like a new level.
Falsifier: simulate from the fitted no-change dependence model or apply BOCPD to prewhitened residuals. Compare false-change frequencies.
15. Gradual drift is a counterexample to the abrupt-break model
Suppose a parameter moves from 0 to 1 over 200 observations. There may be no single true changepoint.
BOCPD may create one or several approximate breaks because piecewise segments are its representational language.
Falsifier: compare with a state-space model that permits continuous parameter evolution. If the smooth model predicts substantially better without discrete resets, “changepoint” is not the only plausible explanation.
16. Multiple plausible run lengths are not a failure
After ambiguous data, the posterior can be multimodal: perhaps one hypothesis places the break 5 observations ago and another 30 observations ago.
Collapsing immediately to the MAP run length hides this uncertainty.
Diagnostic: report posterior mass over meaningful run-length bands and posterior entropy, not only the maximum-probability index.
17. Computational cost grows without approximation
At time t, an exact run-length representation can contain O(t) states. Updating all of them at every step leads to O(t) work per observation and O(T²) cumulative work over a stream of length T, ignoring model-specific constant factors.
That motivates pruning, truncation, capped windows or other approximations in long streams.
But approximation creates a new verification obligation: how much posterior mass was discarded, and did the conclusion depend on it?
18. Pruning can erase a low-probability history that later becomes important
A run-length hypothesis can have small posterior probability today but become more plausible after future observations. If it was permanently pruned, that recovery is impossible.
Falsifier: compare the approximated run with a larger cap or much lower pruning threshold on a representative subset. Measure discarded mass and change in detected break probabilities.
19. Numerical underflow is a silent failure mode
Repeated multiplication of small probabilities can underflow floating-point arithmetic. Implementations therefore commonly operate in log probability and use stable log-sum-exp normalisation.
Falsifier: compare probabilities from standard arithmetic with a log-space reference on long synthetic streams. Zeroed states caused only by numerical underflow are not statistical evidence.
20. Counterexample: missing observations are treated as zero residuals
Replacing a missing observation with zero artificially supports the existing segment if zero is the baseline mean.
Falsifier: model missingness explicitly or skip the likelihood update according to a declared rule, then compare results. Missing data are not evidence of stability.
21. Evidence polarity
Evidence for confidence includes persistent posterior mass on recent breaks after several observations, stability across reasonable hazard priors, good posterior-predictive diagnostics within segments, low discarded pruning mass, low false-break rates on no-change simulations and accurate localisation on injected-break simulations.
Evidence against confidence includes break probabilities driven by one outlier, repeated resets during volatility bursts, strong sensitivity to hazard choice, poor predictive residuals inside claimed segments, heavy pruning, numerical underflow, or better performance from a continuous-drift model.
22. Diagnostics ladder
- Target audit: define the monitored variable or residual.
- Segment-model audit: inspect distribution, tails and dependence.
- Prior audit: document segment-parameter priors.
- Hazard audit: translate H(r) into an interpretable prior segment-length distribution.
- Predictive audit: test one-step predictive calibration.
- Injection audit: simulate known abrupt changes of several sizes.
- No-change audit: measure false resets.
- Outlier audit: inject isolated shocks.
- Gradual-drift audit: test slow parameter changes.
- Dependence audit: simulate autocorrelated baselines.
- Approximation audit: measure pruned/truncated mass.
- Sensitivity audit: vary hazard and prior choices.
- Numerical audit: verify log-space normalisation.
23. BOCPD versus CUSUM
CUSUM accumulates directional evidence for a specified shift and is commonly judged through control limits and average run length. BOCPD maintains a posterior distribution over time since the latest change and combines a segment predictive model with a changepoint hazard.
CUSUM asks, in effect, whether cumulative evidence for a change has crossed a designed threshold. BOCPD asks which possible run lengths are probabilistically plausible after observing the stream.
24. BOCPD versus hidden Markov models
A Hidden Markov Model typically assumes a fixed set of latent states with a transition matrix. States can recur: state A can return after state B.
Standard BOCPD instead focuses on segmentation and time since the latest break. A new segment need not be labelled as one of a small recurring state library.
If recurring economic regimes are the main object, an HMM may be the more natural representation. If unknown break locations and segment duration are central, BOCPD may be more natural.
25. BOCPD versus particle filtering
Particle filters approximate posterior distributions over evolving latent states by weighted particles. They are useful when the hidden state moves continuously or the model is nonlinear/non-Gaussian.
BOCPD’s canonical latent quantity is discrete run length. Extensions can combine changepoint inference with particle methods, but the base page roles remain distinct.
26. Governance: a posterior change is not a causal explanation
An online break can be caused by genuine behaviour, a data-pipeline change, policy revision, product mix, measurement error, seasonality or model misspecification. BOCPD alone cannot identify which mechanism produced the surprise.
As one current model-governance reference, the U.S. Federal Reserve’s SR 26-2 of 17 April 2026 superseded SR 11-7 and SR 21-8 in its stated supervisory context and emphasises risk-based validation, ongoing monitoring and outcome analysis. BOCPD can be one diagnostic among many; it is not a regulatory finding or a replacement for investigation.
27. Connections to the surrounding Bukit Timah Tutor knowledge estate
How Banks Detect Model Drift owns the broad monitoring and redevelopment workflow. BOCPD owns one probabilistic online-break mechanism within that broader problem.
How CUSUM Sequential-Detection Algorithms Find Small Persistent Shifts provides the frequentist cumulative-detection comparison.
How Hidden-Markov Regime Algorithms Infer Financial States owns recurring latent-state inference.
How Particle-Filter Algorithms Track Stochastic Volatility owns sequential Monte Carlo state inference.
How Probability-Integral-Transform Algorithms Test Full Financial Forecast Distributions provides one route for checking whether the segment predictive distribution remains calibrated.
The full public lane is indexed at Finance & Banking Algorithms | Applied Mathematics in Real Financial Systems.
28. What would falsify confidence?
Withdraw confidence if no-change simulations produce frequent false resets, if isolated outliers drive persistent break declarations, if the result disappears under modest hazard changes, if the segment model fails predictive checks, if a continuous-drift model explains the data substantially better, or if pruning/truncation discards enough mass to change the posterior conclusion.
29. Verification protocol
Preserve the observation stream, segment likelihood, prior hyperparameters, hazard function, all sufficient-statistic updates, run-length posterior at each step, normalisation constants, pruning policy, discarded probability mass, random seeds for simulations and code version.
Verification should include hand-checkable short sequences, no-change simulations, injected abrupt changes, outlier contamination, gradual drift, autocorrelation, missing data and long-stream numerical stability.
30. Update triggers
Revalidate when the monitored variable changes definition, the upstream forecasting model changes, residual tails or dependence change, the expected regime duration changes, data frequency changes, new seasonality appears, pruning thresholds are altered, missingness changes, or empirical false-change and detection-delay rates move away from the validated operating range.
31. Primary and high-quality references
- Ryan Prescott Adams and David J. C. MacKay, Bayesian Online Changepoint Detection, 2007.
- Reza Habibi, Bayesian Online Change Point Detection in Finance, Financial Internet Quarterly.
- Matias Altamirano, François-Xavier Briol and Jeremias Knoblauch, Robust and Scalable Bayesian Online Changepoint Detection, Proceedings of Machine Learning Research, 2023.
- Board of Governors of the Federal Reserve System, SR 26-2: Revised Guidance on Model Risk Management, 17 April 2026.
Educational boundary: a high posterior probability of a recent changepoint means that a segmented probabilistic model finds a recent reset plausible. It does not identify cause, predict market direction, establish wrongdoing or justify an automatic financial action.