Forecasting-oriented changepoint detection

Introduction. Structural changes are a major source of forecast degradation in time series analysis. Classical changepoint methods typically identify piecewise-constant or piecewise-linear regimes, but such descriptive segmentations do not necessarily correspond to changes in predictive behavior. In many applications, a changepoint does not reflect a modification of the underlying data-generating process (DGP), but rather a loss of predictive homogeneity in the forecasting model. This work introduces a forecasting-oriented segmentation framework that defines regime boundaries according to changes in predictive performance.

Methods. The proposed approach partitions a time series into segments within which a forecasting model remains locally adequate. Candidate segmentations are evaluated through predictive loss functions and information-theoretic criteria (e.g., AIC), rather than approximation error. This formulation is more general than parameter drift-based approaches, which detect instability through changes in estimated parameters, Fisher information, or local convergence properties. While parameter drift can be used as a diagnostic to validate predictive segmentation, it is not required by the framework. The objective function remains additive across segments, but predictive loss functions may not satisfy sub- or super-additivity, which affects the efficiency of classical optimization strategies. This perspective aligns with the emerging view that forecasting performance, rather than structural approximation, should guide regime identification in nonstationary environments.

Results. Experiments on synthetic and real-world datasets show that predictive model-based segmentation identifies regime boundaries aligned with shifts in forecasting adequacy, even when the DGP remains stable. Across multiple model classes, the proposed framework yields improved forecast accuracy compared to approximation-based and parameter drift-based methods.

Conclusions. Recasting changepoint detection as a forecasting-driven segmentation problem provides a principled way to integrate structural change detection, model adequacy assessment, and forecast generation. This perspective generalizes parameter-instability approaches and offers a flexible foundation for predictive regime identification.