lstm change-point detection

This study addresses the automatic detection and precise localization of change points in large-scale multivariate time series exhibiting dynamic evolution. We propose a two-stage change point detection method based on feedforward neural networks (FNNs). The method innovatively integrates FNNs into the change point detection framework, incorporating piecewise training, sliding-window error evaluation, and an online recalibration mechanism, along with a dedicated error calibration strategy to ensure estimation consistency under temporal dependence. Theoretically, we establish consistency of the change point localization estimator. Empirically, the method achieves high accuracy in estimating both the number and locations of change points on both synthetic and real-world datasets. Moreover, it supports practical, data-driven selection of hyperparameters, enhancing its applicability in real-world scenarios.

Addressing the challenges of real-time changepoint detection in semi-structured time-series data (e.g., sensor and video streams)—namely high latency, low accuracy, and difficulty in detecting abrupt disorders—this paper proposes an end-to-end differentiable changepoint detection (CPD) framework. We design a principled, differentiable CPD loss function that jointly optimizes detection latency and false positive rate, integrated with deep representation learning for unified optimization. To support rigorous evaluation, we introduce and publicly release the first video benchmark dataset featuring precise, frame-level disorder annotations. On explosion event detection in videos, our method achieves an F1 score of 0.53, substantially outperforming state-of-the-art baselines (0.31 and 0.35). Furthermore, comprehensive experiments on synthetic sequences and real-world sensor data demonstrate strong generalization and robustness across diverse modalities and noise conditions.

This work investigates the zero-shot and few-shot time-series anomaly detection capabilities of large language models (LLMs). Methodologically, it conducts multi-model comparative experiments (including Llama, Qwen, and GPT series), employs time-series image-based encoding, designs a hypothesis-driven controllable evaluation framework, and applies systematic prompt engineering. The study yields four counterintuitive findings: (1) LLMs perform significantly better when processing time-series as images rather than text; (2) explicit chain-of-thought or reasoning prompts yield no statistically significant improvement; (3) repetition bias and arithmetic reasoning are not primary mechanisms underlying anomaly recognition; and (4) architectural differences lead to substantial performance variation. These results challenge prevailing assumptions in the field and empirically demonstrate that LLMs possess foundational—but non-trivial—time-series anomaly detection capabilities. To foster reproducibility and further research, the authors open-source both the implementation code and a dedicated benchmark dataset, AnomLLM.

This paper addresses poor model reproducibility and insufficient open-source implementations in time-series forecasting by proposing a lightweight, fully reproducible LSTM/GRU modeling paradigm. Methodologically, it constructs univariate sequence samples via sliding windows and evaluates performance using two metrics—RMSE and directional accuracy (DA)—on both synthetic activity data (Activities) and real-world financial data (BSE BANKEX). A key finding is that effective training requires only a single time series exhibiting repetitive patterns, without complex preprocessing or large-scale datasets. Experiments show that the proposed implementation significantly outperforms the “repeat last value” baseline for 1-step and 20-step predictions on Activities, while achieving comparable performance on BSE BANKEX. All code, datasets, and complete experimental configurations are publicly released to ensure full reproducibility and out-of-the-box usability.

This paper addresses sequential change-point detection in compositional time series with exogenous variables (e.g., COVID-19 positivity rates). Methodologically, it introduces a generalized Beta AR(1) dynamic model that accommodates exogenous covariates; develops a partial maximum likelihood estimation (MLE) framework and rigorously establishes its consistency and asymptotic normality; derives necessary and sufficient conditions for strict stationarity and geometric ergodicity of the model; and constructs a parametric CUSUM-type sequential test statistic. The key contribution is the first theoretically grounded, interpretable, online change-point detection framework specifically designed for compositional time series with exogenous inputs—unifying dynamic modeling and change-point inference under provable statistical guarantees. Empirical evaluation on real-world COVID-19 data demonstrates high detection sensitivity and practical utility.

This study addresses the challenges of production disruptions and high maintenance costs caused by sudden failures in industrial hydraulic pumps by proposing an unsupervised early fault detection method that relies solely on normal operational data. The approach compares a feedforward autoencoder, which processes single-frame sensor snapshots, with an LSTM-based autoencoder that models short-term temporal windows. Both models are trained exclusively on 52-channel, minute-level sensor logs without any fault examples. Evaluated on an independent test set containing seven annotated fault intervals, both architectures achieve highly reliable detection performance. The results validate the effectiveness of temporal modeling for industrial anomaly detection and demonstrate the practical feasibility and real-world applicability of unsupervised methods in industrial settings.

This work addresses the problem of efficient online change-point detection in both univariate and multivariate data streams by introducing a novel method grounded in the Focus algorithm family. Leveraging the generalized likelihood ratio test, the approach enables exact detection of a single change point without requiring approximations. By exploiting the relationship between candidate change-point locations and the geometric structure of the data, it achieves a computational complexity of approximately $\log(n)^d$ per iteration. Notably, this is the first method to support exponential-family models, nonparametric settings, and autoregressive data under no approximation assumptions, integrating natural exponential-family modeling, empirical cumulative distribution functions, and geometric optimization techniques. The accompanying R/Python software package substantially enhances the efficiency and applicability of change-point detection in high-dimensional streaming data.

This study addresses the inadequacy of traditional actuarial methods in estimating loss reserves under increasing climate-driven catastrophe frequency, which violates the stability assumptions underlying conventional approaches and leads to significant estimation bias. To overcome this limitation, we propose the first application of Long Short-Term Memory (LSTM) neural networks to insurance reserving, integrating climate covariates—such as NOAA hurricane intensity indices and sea surface temperature—to detect and adapt to structural shifts in over a decade of regulatory loss triangles from Florida and Louisiana. We develop a probabilistic framework for this climate-augmented LSTM model, providing formal performance guarantees, and benchmark it against standard methods including Chain Ladder, Bornhuetter–Ferguson, and Cape Cod. Empirical results demonstrate that our approach improves reserve estimation accuracy by 15%–20% in catastrophe years, effectively mitigating challenges posed by the sparsity of extreme events.