Record-Driven Blockwise Anomaly Detection in Weakly Dependent Time Series

We consider a real-valued time series {X_t}t≥1 and aim at detecting changepoints understood as persistent departures from a reference local regime. We propose adistribution-free method that combines (i) a local power statistic {Pt} derived from localasymptotic normality (LAN) testing framework, (ii) blockwise aggregation of exceedanceevents, and (iii) a principled calibration of the tolerance parameter d using record theory(under the independent and identically distributed (i.i.d.) , linear-drift (LDM), and YangNevzorov models). Under β-mixing, block exceedance counts admit a binomial approximation, yielding tractable decision rules that separate normal behavior, transient anomalies,and genuine changepoints. Simulations illustrate the practical performance and robustnessof the approach