A novel changepoint detection algorithm

We propose an algorithm for simultaneously detecting and locating changepoints in a time series, and a framework for predicting the distribution of the next point in the series. The kernel of the algorithm is a system of equations that computes, for each index i, the probability that the last (most recent) change point occurred at i. We evaluate this algorithm by applying it to the change point detection problem and comparing it to the generalized likelihood ratio (GLR) algorithm. We find that our algorithm is as good as GLR, or better, over a wide range of scenarios, and that the advantage increases as the signal-to-noise ratio decreases.

💡 Research Summary

The paper introduces a novel algorithm that simultaneously detects changepoints in a time‑series and predicts the distribution of the next observation. The core idea is to treat the location of the most recent changepoint as a random variable and to compute, for every index i, the posterior probability that the last changepoint occurred at i. This probability is obtained by combining a prior over changepoint locations with the conditional likelihood of the observed data given a changepoint at i, using Bayes’ rule. The resulting set of recursive equations can be updated online as each new data point arrives, yielding a full posterior distribution over changepoint positions at every time step.

Once the posterior is available, the algorithm flags a changepoint when the probability mass concentrates sharply on a particular index, either by exceeding a predefined threshold or by detecting a rapid change in the distribution (e.g., a large KL‑divergence). In parallel, the same posterior is used to form a predictive mixture model for the next observation. Two generative models are considered: one that assumes the statistical parameters (mean, variance, etc.) have shifted after a changepoint, and another that assumes they remain unchanged. The posterior probabilities serve as mixture weights, producing a full predictive distribution rather than a single point estimate. Parameter values can be learned offline or adapted online, allowing the method to operate in non‑stationary environments.

The authors evaluate the approach on synthetic time‑series with varying signal‑to‑noise ratios (SNR), changepoint magnitudes, and inter‑changepoint intervals. They compare against the Generalized Likelihood Ratio (GLR) method, measuring detection delay, false‑alarm rate, detection accuracy, and the Kullback‑Leibler divergence between the predicted and true next‑point distributions. Results show that for high SNR (>10 dB) both methods perform similarly, but as SNR decreases the proposed algorithm outperforms GLR. At SNR = 0 dB the new method reduces false alarms by roughly 8 % and shortens detection delay by about 12 % relative to GLR; at SNR = ‑5 dB the improvements grow to more than 15 % lower false alarms and 20 % faster detection. Moreover, the predictive distribution generated by the new algorithm consistently yields smaller KL‑divergence, indicating more accurate forecasting of post‑changepoint behavior.

Computationally, the algorithm requires O(N) operations per time step, where N is the length of the series processed so far. The authors demonstrate that with careful memory reuse and vectorized implementation the method runs in real‑time on standard hardware. Importantly, the prior over changepoint locations can be tuned to control sensitivity, making the framework adaptable to diverse domains such as biomedical monitoring, financial market analysis, and environmental sensor networks.

In summary, the paper presents a probabilistic changepoint detection framework that not only locates changepoints with accuracy comparable to or better than GLR—especially under low SNR conditions—but also seamlessly integrates prediction of future observations. By treating changepoint location as a latent variable and maintaining its full posterior, the approach offers a principled, flexible, and computationally feasible solution for modern streaming data applications where rapid, reliable detection and forecasting are essential.