Speaker
Description
Change-point detection in time series is complicated by the absence of reliable ground-truth labels of regime changes. Real data often do not provide verified transition moments, while expert annotation is subjective and difficult to
scale. Therefore, synthetic benchmarks are useful for evaluating both machine-learning models and the validity of the supervised detection problem itself.
This work proposes a controlled synthetic benchmark for change-point detection. The benchmark is based on a stochastic generator with a bistable structure and several types of coloured noise. Since transition events are defined by construction, the setup allows models to be compared under reproducible conditions.
A unified experimental pipeline was implemented for data generation, transition labelling, sliding-window construction,model training, and evaluation. The pipeline supports recurrent neural networks, transformer-based models, Mambastyle
architectures, and gradient boosting. Detailed grid experiments were conducted using a GRU detector as a computationally efficient representative model.
The principal results are summarised in three experimental blocks. First, the dwell-time statistics and regime classification block analysed the empirical distribution of inter-transition intervals. The generator was evaluated
on 100 configurations over noise colour, noise intensity, and integration time step. Out of these configurations, 82 were valid for dwell-time fitting. Among them, 63 cells, or 76.8%, corresponded to an intermediate lognormal-like
regime; 17 cells, or 20.7%, corresponded to a normal/Kramers-like regime; and only 2 cells, or 2.4%, were classified as heavy-tailed.
Second, the horizon and relaxation block studied how the prediction horizon and relaxation window affect transition prediction quality using the GRU detector. In this supervised grid of 210 configurations, the best GRU performance
reached ROC-AUC 0.9947 and was obtained under blue noise.
