Optimal ensemble size for Ensemble-SCAN

Determine the optimal number of base SCAN detectors in the Ensemble-SCAN procedure, balancing stability across window sizes against the additional computational cost of including more detectors.

Background

Ensemble-SCAN aggregates detections from a finite collection of SCAN detectors using different window sizes. The supplementary material explains that increasing the ensemble size can improve stability across window choices but also increases computational cost, while the theoretical consistency argument treats the ensemble size as fixed.

The paper does not identify a theoretically or empirically optimal ensemble size. Resolving this question would clarify how the number of detectors should scale with the time-series length, available computational resources, and desired detection stability.

References

Although the literature does not provide an explicit optimal choice for the number of SCAN detectors in the ensemble, increasing this number excessively is not necessarily beneficial because it increases computational cost.

— SCAN: Sequentially Detecting Change-points via Adaptive Nonparametric Inference  (2608.28110 - Prabashwara et al., 28 Aug 2026) in Section “Computational Complexity and Parallel Implementation”

We suspect this might be caused by how the framework works: the evolutionary framework often asks to merge different ideas together, and (by construction) it also presents diverse programs as context at each evolution step, so incremental additions to an already competitive program tend to be kept even when they grow a larger, less coherent ensemble rather than a fresh single principle. A single run per regime cannot pin down that mechanism, but it is a plausible reason the shared prompt did not prevent this kind of heuristic accretion.

— Have an LLM Write Your Anomaly Detector: Autonomous Discovery of Compact, Interpretable Detectors for Time Series  (2610.01223 - Berghaus, 1 Oct 2026) in Section 5.2, “The search procedure matters less, but shapes what gets built”