TimeRCD: Foundation Model for Zero-Shot Time Series Anomaly Detection
A new foundation model for time series anomaly detection (TSAD), named TimeRCD, has been unveiled by researchers. This model distinguishes itself by functioning in a zero-shot manner, unlike traditional models that depend on reconstruction-error scoring. Instead, TimeRCD employs Relative Context Discrepancy (RCD), a pre-training strategy that assesses a query pattern against its contextual surroundings to identify anomalies. Utilizing a standard Transformer architecture, this model derives normality from the input context rather than relying on static global patterns. Additionally, the research team developed a large synthetic dataset featuring context-dependent anomaly labels to facilitate supervised pre-training for RCD. The findings are published on arXiv with the identifier 2509.21190.
Key facts
- TimeRCD is a foundation model for zero-shot time series anomaly detection.
- It uses Relative Context Discrepancy (RCD) pre-training paradigm.
- RCD compares a query pattern with its surrounding context.
- The model uses a standard Transformer architecture.
- It infers normality from input context rather than fixed global patterns.
- A large-scale synthetic corpus with context-dependent anomaly labels was constructed.
- The paper is on arXiv with identifier 2509.21190.
- Existing models often rely on reconstruction-error scoring.
Entities
Institutions
- arXiv