RAMSeS: Robust and Adaptive Model Selection for Time-Series Anomaly Detection Algorithms

Mohamed Abdelmaksoud; Sheng Ding; Andrey Morozov; Ziawasch Abedjan

cs.DBFeb 25, 2026

RAMSeS: Robust and Adaptive Model Selection for Time-Series Anomaly Detection Algorithms

Authors: Mohamed Abdelmaksoud, Sheng Ding, Andrey Morozov, Ziawasch Abedjan

Organizations: BIFOLD & TU Berlin Berlin, Germany · University of Stuttgart Stuttgart, Germany

Abstract

Time-series data vary widely across domains, making a universal anomaly detector impractical. Methods that perform well on one dataset often fail to transfer because what counts as an anomaly is context dependent. The key challenge is to design a method that performs well in specific contexts while remaining adaptable across domains with varying data complexities. We present the Robust and Adaptive Model Selection for Time-Series Anomaly Detection RAMSeS framework. RAMSeS comprises two branches: (i) a stacking ensemble optimized with a genetic algorithm to leverage complementary detectors. (ii) An adaptive model-selection branch identifies the best single detector using techniques including Thompson sampling, robustness testing with generative adversarial networks, and Monte Carlo simulations. This dual strategy exploits the collective strength of multiple models and adapts to dataset-specific characteristics. We evaluate RAMSeS and show that it outperforms prior methods on F1.

Figures & tables

Fig. 1: F1 score for six anomaly detection models from statistical and neural-network families across four domains (Prices, Medical, Industry, and Environment) [ 21 , 29 , 30 , 31 ] . Red vertical lines indicate spike anomalies. The triangle marks the best-performing model in each domain.

Fig. 3: SKAB 1-1 time series

Fig. 4: Stacking ensemble via a genetic algorithm.

Fig. 5: Linear Thompson sampling with $\varepsilon$ -greedy for model selection.

Fig. 6: Continuous online real-time model selection for anomaly detection.

Base models	Category	Dim	Brief Description
AutoEncoder [ 82 ]	NN	U&M	Reconstruction error
RNN [ 83 ]	NN	U&M	Vanilla recurrent prediction error
DGHL [ 84 ]	NN	M	Gen-ConvNet with latent factor
LSTMVAE [ 25 ]	NN	U&M	LSTM-based VAE
Donut [ 85 ]	NN	U&M	VAE likelihood scoring
OmniAnomaly [ 86 ]	NN	U&M	Stochastic RNN likelihood

TABLE I: Base models grouped by family: Neural Networks (NN), Statistical (Stat) or Foundation Models (FM) and applicability (U/M).

Entity	Parameter	Value	Rationale
GA-Ens	Population size ( $P$ )	$\leq$ 20	F1 saturation ( VI-H )
	Number of Generations ( $G$ )	$\leq$ 20	Cost-effective ( VI-H )
	Mutation rate ( $\mu$ )	$\leq$ 0.2	Sustains diversity
	Meta-learner	RF	Nonlinear aggregation
LinTS	Windows	50	Posterior convergence
	Epsilon ( $\epsilon_{0}$ )	0.2	Early exploration

TABLE II: RAMSeS Hyperparameters for Reproducibility

Method	UCR (250 TS)			SMD (27 TS)			SKAB (18 TS)
Method	F1	T	M	F1	T	M	F1	T	M
TSB	0.45	5012	534	0.55	4603	825	0.68	368	702
UMS	0.66	2661	2580	0.80	1520	1796	0.86	546	1312
MSAD	0.07	14918	639	0.13	260	608	0.54	74	286
AutoTSAD	Timeout			Not applicable			Not applicable
RAMSeS	0.83	1704	785	0.89	586	618	0.86	647	512

TABLE III: Average System-level performance: RAMSeS vs. automated TSAD baselines across all datasets and Time series (TS). Time (T) in seconds, Memory (M) in MB. Best in bold . Timeout is 48 hours per time series.

Dataset	GA	LinTS	GAN	SBA	MC	Single Branch	System Memory
UCR	75	1445	71	39	73	1628	785
SMD	280	173	38	30	64	306	618
SKAB	134	337	75	42	88	513	512

TABLE IV: Average end-to-end runtime (seconds) and memory (MB) per entity.

Fig. 7: Online phase avg. F1 & overhead on SMD for RAMSeS & Baselines

Fig. 8: Scalability: F1 and latency vs. candidate pool size for SKAB dataset.

Fig. 9: Window size sensitivity on SKAB dataset.

Dataset	Domain	N	GA-Ens	Single Sel.	Final Sel.
			F1± $\sigma$	F1± $\sigma$	F1± $\sigma$
UCR-Med	Healthcare	68	0.77±0.06	0.80±0.09	0.82±0.07
UCR-Env	Environment	13	0.62±0.01	0.80±0.09	0.81±0.04
UCR-Ind	Industry	33	0.72±0.07	0.80±0.09	0.82±0.07
SMD	IT	25	0.88±0.07	0.87±0.05	0.89±0.05
SKAB	Industry	15	0.78±0.04	0.55±0.30	0.80±0.07

TABLE V: Branch average performance comparison based on F1-score.

Entity	P 10			P 50			P 100
	G 100	1000	10000	100	1000	10000	100	1000	10000
SKAB 1-1	0.87	0.87	0.87	0.87	0.87	0.87	0.87	0.87	0.87
SKAB 1-2	0.87	0.87	087	0.87	0.87	0.87	0.87	0.87	0.87
SMD 3-4	0.77	0.76	0.76	0.76	0.76	0.76	0.76	0.76	0.76
SMD 3-10	0.71	0.71	0.71	0.71	0.71	0.71	0.71	0.71	0.71
011_UCR	0.77	0.77	0.77	0.77	0.77	0.77	0.77	0.77	0.77

TABLE VI: F1 across

P

sizes and

G

. Meta-learner= SVM and

\mu

= 0.2.

Table 16

Fig. 10: Runtime–accuracy of the GA-based stacking ensemble on SKAB1_1 . (A) runtime vs. work ( $P{\times}G$ ), $R^{2}{=}0.931$ . (B) runtime vs. $\mu$ across meta-learners. (C) runtime vs. meta-learner at three workloads. (D) $F_{1}$ vs. work, $R^{2}{=}0.540$ . (E) $F_{1}$ vs. $\mu$ . (F) $F_{1}$ vs. meta-learner. Points are medians over 3 runs.

Table 18

Explore similar work

Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection
Oct 14, 2026Christos Panourgias, Roberto Stanzione, Adrien Petralia +2Model Selection Time-Series Anomaly Detection
Université Paris Cité, LIPADE · École Normale Supérieure, Inria
ChronosAD: Leveraging Time Series Foundation Models for Accurate Anomaly Detection
May 31, 2026Uzair Khan, Luigi Capogrosso, Francesco Biondani +4Industrial Anomaly Detection Time-Series Anomaly Detection
University of Verona · Interdisciplinary Transformation University of Austria · ETH Zurich
OneShot: Label-Efficient Detector Weighting for Time Series Anomaly Detection
Oct 12, 2026Seunghan Lee, Jun Seo, Jaehoon Lee +9Time-Series Anomaly Detection Anomaly Detection
LG AI Research

Oct 14, 2026J/K move · Enter open · S save

Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection

Christos Panourgias, Roberto Stanzione, Adrien Petralia, Themis Palpanas+1

Université Paris Cité, LIPADE · École Normale Supérieure, Inria

Model Selection Time-Series Anomaly Detection OOD Generalization

Open paper AskPDF arXiv

Anomaly detection is a fundamental task for time-series analytics, with important implications for the downstream performance of many applications. Despite the large number of anomaly detection methods proposed in the literature, recent benchmark studies have shown that no single detector performs best across highly heterogeneous time series. Therefore, a practical and scalable solution is to develop a model-selection method that, for a given time series, selects the anomaly detector most likely to perform well. Nevertheless, the model selection approaches proposed in the literature suffer from a significant drop in accuracy when applied in Out-of-Distribution (OOD) settings. In this paper, we tackle the aforementioned limitation and propose RAMSAD, a retrieval-based framework for model selection in time-series anomaly detection. Given a new time series, our method queries a knowledge base of previously observed instances, retrieves the top-k most similar series, and transfers detector recommendations from their performance profiles. The framework can operate with standard similarity measures as well as embedding-based representations. Our results show that RAMSAD outperforms existing baselines in In-Distribution (ID) settings and is the most robust selector in OOD, where, when selecting a small subset of detectors, it matches or exceeds ensembles that run the entire detector pool, at a fraction of their cost. Overall, we demonstrate that similarity-based retrieval constitutes a strong and efficient foundation for model selection in time-series anomaly detection.

Abstract

Explore similar work

Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection

ChronosAD: Leveraging Time Series Foundation Models for Accurate Anomaly Detection

OneShot: Label-Efficient Detector Weighting for Time Series Anomaly Detection