Anomaly detection is a core task in time-series analytics, and the choice of detection method can significantly affect downstream performance. A new paper on arXiv, 'Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection,' tackles the problem of model selection in this setting.

The authors propose a retrieval-based approach: instead of evaluating all candidate models on a new time series, they search for similar training examples and use those to select the most appropriate model. This 'retrieval over training' strategy is designed to reduce computational cost while maintaining accuracy.

The abstract notes that many anomaly detection methods exist, but it does not yet provide experimental results or comparisons in the visible portion. The full paper presumably details the retrieval mechanism and its performance.