Time-Series Paper Digest: Nov 1‑7 2025 Highlights

DoFlow: Causal Generative Flows for Interventional and Counterfactual Time‑Series Prediction

Paper link: http://arxiv.org/pdf/2511.02137v1

Authors: Dongze Wu, Feng Qiu, Yao Xie

Abstract: Time‑series forecasting increasingly requires not only accurate observational predictions but also causal predictions for interventions and counterfactual queries in multivariate systems. DoFlow is a flow‑based generative model built on a causal directed acyclic graph (DAG). It uses continuous‑normalizing flows (CNFs) for natural encoding and decoding, delivering consistent observational, interventional, and counterfactual forecasts. Under certain assumptions the authors also provide theoretical results for counterfactual recovery. Beyond prediction, DoFlow yields explicit likelihoods of future trajectories, enabling principled anomaly detection. Experiments on synthetic datasets with varied causal DAGs and real‑world hydro‑electric and cancer‑treatment time‑series demonstrate accurate system‑level observational forecasts, successful causal inference for interventions and counterfactuals, and effective anomaly detection.

Forecast2Anomaly (F2A): Adapting Multivariate Time Series Foundation Models for Anomaly Prediction

Paper link: http://arxiv.org/pdf/2511.03149v1

Authors: Atif Hassan, Tarun Kumar, Ashish Mishra, Sergey Serebryakov, Satish Kumar Mopur, Phanidhar Koganti, Murthy Chelankuri, Ramanagopal Vogety, Suparna Bhattacharya, Martin Foltin

Abstract: Predicting anomalies in multivariate time‑series from dynamic, complex real‑world systems is critical for preventing major failures and reducing operational costs. Existing methods are system‑specific and fail to generalize to evolving anomaly patterns. Pre‑trained time‑series foundation models (TSFMs) have shown strong generalization and zero‑shot prediction abilities, yet their potential for anomaly prediction remains untapped because anomaly detection differs fundamentally from normal‑behavior forecasting. F2A introduces two innovations: (1) a joint prediction‑anomaly loss that fine‑tunes TSFMs to accurately forecast future signals even at anomalous timestamps, and (2) a retrieval‑augmented generation (RAG) module that retrieves historically relevant horizons and conditions the forecast on them. The RAG component dynamically adapts to distribution shift at inference time, allowing F2A to track evolving anomalies without model updates. Extensive experiments on 16 datasets and multiple TSFM backbones show that F2A consistently outperforms state‑of‑the‑art methods, offering a scalable zero‑shot anomaly prediction solution for real‑world applications.

ForecastGAN: A Decomposition‑Based Adversarial Framework for Multi‑Horizon Time Series Forecasting

Paper link: http://arxiv.org/pdf/2511.04445v1

Authors: Syeda Sitara Wishal Fatima, Afshin Rahimi

Abstract: Time‑series forecasting is vital across domains such as finance and supply‑chain management. Existing methods struggle with multi‑horizon forecasts, especially when short‑term accuracy suffers despite strong long‑term performance of Transformer models. ForecastGAN proposes a novel decomposition‑based adversarial framework that addresses these limitations. It consists of three integrated modules: (1) a decomposition module that extracts seasonal and trend components, (2) a model‑selection module that identifies the optimal neural network configuration based on forecast horizon, and (3) an adversarial training module that employs a conditional generative adversarial network to enhance robustness. Unlike traditional approaches, ForecastGAN seamlessly incorporates both numerical and categorical features. Evaluation on eleven benchmark multivariate time‑series datasets covering various horizons shows that ForecastGAN consistently surpasses state‑of‑the‑art Transformers on short‑term forecasts while remaining competitive on long‑term horizons, all without extensive hyper‑parameter tuning.