Skip to content

Causally-Constrained Probabilistic Forecasting for Time-Series Anomaly Detection

Home / Papers / Causally-Constrained Probabilistic Forecasting for Time-Series Anomaly Detection

Causally-Constrained Probabilistic Forecasting for Time-Series Anomaly Detection

Authors: Pooyan Khosravinia, João Gama, Bruno Veloso Date: 2026-04-20 Paper ID: openalex:2604.17998

Summary

This paper introduces a Causally Guided Transformer (CGT) for multivariate time-series anomaly detection that incorporates time-lagged causal graph priors to bridge the gap between correlation-based deep learning and causal interpretability. The model utilizes a hard parent mask to restrict predictions to causal precursors and a shadow auxiliary path with safety-gated blending to selectively incorporate non-causal correlational data. Evaluation on the ASD and SMD benchmarks demonstrates that this causally-constrained approach outperforms existing models in both detection accuracy and root-cause localization precision.

Key Contributions

  • Introduces a Causally Guided Transformer (CGT) that uses time-lagged causal graph priors to constrain multivariate time-series forecasting.
  • Implements a safety-gated blending mechanism to isolate causal predictions from residual correlational information.
  • Achieves SOTA anomaly detection on ASD (96.19% F1) and SMD (95.32% F1) datasets while providing improved root-cause localization through counterfactual clamping.

Open Questions & Future Work

Key Concepts

  • causally-guided-transformer: A transformer architecture that restricts predictive paths to graph-supported causal precursors via hard parent masking.

Archivist Review

Approved the Causally Guided Transformer as a distinct architectural paradigm for incorporating structural priors. The dataset ASD was rejected as a standard benchmark, while SMD was kept to track core industrial evaluation. The open question was reframed to focus on the bottleneck of joint end-to-end causal learning.

Approved Concepts

  • Causally Guided Transformer: It provides a reusable framework for incorporating structural inductive biases into transformer-based forecasting via hard parent masking.

Approved Open Questions

  • Joint Causal Discovery Bottleneck: This question addresses the core tension between static causal priors and dynamic real-world environments, which is essential for advancing robust, interpretable forecasting.

Rejected Candidates

  • [dataset] ASD (asd) - not_novel: Routine industrial anomaly detection dataset.

Datasets

Metadata & Links

url
https://arxiv.org/abs/2604.17998
paper_id
2604.17998
paper_source
openalex
domain
time-series
tags
anomaly-detectionforecastingcausalitymultivariate-time-seriestransformer
architectures
datasets
smd
concept_slugs
causally-guided-transformer
dataset_slugs
smd
skill
TimeSeriesSkill
created_at
2026-04-23T05:47:00Z