An Interpretable Generative Framework for Anomaly Detection in High-Dimensional Financial Time Series

Waldyn G Martinez

stat.MLMar 9, 2026

An Interpretable Generative Framework for Anomaly Detection in High-Dimensional Financial Time Series

Authors: Waldyn G Martinez

Organizations: Department of Information Systems & Analytics Farmer School of Business Miami University Oxford, OH 45056

Abstract

Detecting structural instability and anomalies in high-dimensional financial time series is challenging due to complex temporal dependence and evolving cross-sectional structure. We propose ReGEN-TAD, an interpretable generative framework that integrates modern machine learning with econometric diagnostics for anomaly detection. The model combines joint forecasting and reconstruction within a refined convolutional--transformer architecture and aggregates complementary signals capturing predictive inconsistency, reconstruction degradation, latent distortion, and volatility shifts. Robust calibration yields a unified anomaly score without labeled data. Experiments on synthetic and financial panels demonstrate improved robustness to structured deviations while enabling economically coherent factor-level attribution.

Figures & tables

Panel of Figure 1 — Figure 1: Rolling-window construction. The model ingests past $L$ observations $\mathbf{X}_{t}$ and jointly predicts the future block $\mathbf{F}_{t}$ over horizon $H$ .

Panel of Figure 2 — Figure 2: Temporal Convolution over the Input Window. A convolutional filter of width three is applied to the first three time steps of $\mathbf{X}_{t}$ , operating jointly across all $p$ features. The shaded region represents the receptive field used to produce a feature vector $\mathbf{h}_{t-L+3}\in\mathbb{R}^{d}$ .

Panel of Figure 3 — Figure 3: Positional encoding. The embedding matrix $\mathbf{H}_{t}$ is augmented by the deterministic positional encoding matrix $\mathbf{P}$ through element-wise addition, producing $\tilde{\mathbf{H}}_{t}$ while preserving dimensionality.

Figure 4: Illustration of ReGEN-TAD across six synthetic financial regimes: Bull Market, Mean Shift, Volatility Spike, Flash Crash, Regime Switch, and Liquidity Dryup. Top panels show multivariate price trajectories (sample of first 5 series) with anomaly region shaded. Bottom panels show corresponding ReGEN-TAD anomaly scores and predicted anomalies.

Figure 5: Overall architecture of ReGEN-TAD. The generative backbone transforms input windows through convolutional, positional, attention-based, and recurrent encoding to produce a latent representation. Forecasting and reconstruction heads generate diagnostic signals, which are aggregated in the decision function using robust scaling and smoothing before applying rank- or threshold-based anomaly selection.

Figure 6: Factor-level attribution for sector-specific synthetic anomalies detected by ReGEN-TAD. Each panel corresponds to a distinct anomaly type injected into a single economic sector (financials, technology, industrials, and defensive assets). For each anomaly, the top subpanel displays the normalized ReGEN-TAD anomaly score over time, with detected anomalous windows highlighted, while the bottom subpanel reports the top contributing factors for the explained window based on standardized deviation from the baseline and reconstruction sensitivity. The results illustrate ReGEN-TAD’s ability to localize anomalies temporally and attribute them to economically relevant factors.

Model	F1	Precision	Recall	AUROC	FPR	Runtime (s)
ReGEN-TAD	0.6239	0.8340	0.5954	0.9477	0.0092	24.5519
GARCH	0.5945	0.6635	0.6261	0.8432	0.0228	0.0152
OLS	0.5457	0.6978	0.5175	0.9549	0.0042	3.0716
RRR	0.5141	0.6519	0.4931	0.8860	0.0038	3.6446
TimeGPT	0.4949	0.7437	0.4039	0.6970	0.0099	3.6237
TranAD	0.4395	0.6054	0.4614	0.7218	0.0148	8.0562

Table 1: Overall Average Performance Metrics across all Anomaly Types and Contamination Rates from Synthetic Data. Models are ranked by F1 Score.

Model	F1	Precision	Recall	AUROC	FPR	Runtime (s)
ReGEN-TAD	0.7717	0.9684	0.6427	0.9704	0.0018	12.3828
OLS	0.6638	0.7636	0.6499	0.9747	0.0044	4.4457
RRR	0.6165	0.7197	0.6048	0.9364	0.0040	5.0803
TranAD	0.5612	0.7010	0.5721	0.7977	0.0146	4.4906
TimeGPT	0.5307	0.7795	0.4348	0.7132	0.0084	3.8887
GARCH	0.4905	0.5690	0.4874	0.7840	0.0292	0.0117

Table 4: Overall Average Performance Metrics across all Anomaly Types and Contamination Rates. Models are ranked by F1 Score.

Figure 7: F1 score across prediction horizons and anomaly types. Short- to medium-term horizons yield stronger overall detection performance.

Model	Metric	Bear	Bull	Cont.	Corr. Brk	Fat Tail	Flash	Liq. Dry	Micro.	Mom. Crsh	Regime	Sector	Trend	Vol. Spk
ReGEN-TAD	F1	0.7831	0.7831	0.7944	0.8056	0.8056	0.5287	0.8056	0.7843	0.8056	0.8056	0.8056	0.7859	0.7859
	FPR	0.0002	0.0002	0.0008	0.0000	0.0000	0.0189	0.0000	0.0015	0.0000	0.0000	0.0000	0.0000	0.0000
	Precision	0.9967	0.9967	0.9862	1.0000	1.0000	0.6723	1.0000	0.9739	1.0000	1.0000	1.0000	1.0000	1.0000
	Recall	0.6464	0.6464	0.6660	0.6753	0.6753	0.4367	0.6753	0.6573	0.6753	0.6753	0.6753	0.6488	0.6488
	Runtime (s)	13.9892	14.0268	10.5484	10.5476	10.6600	14.3892	17.3604	10.4007	10.6488	10.4999	10.4864	14.2313	14.3094
OLS	F1	0.6901	0.6901	0.8955	0.9085	0.9204	0.5073	0.9477	0.5478	0.9468	0.9419	0.8983	0.5059	0.9180

Table 5: Performance metrics by Model and Anomaly Type (Low Contamination

\gamma\in\{0.01,0.03,0.05\}

). Models are ordered by overall F1 score.

Figure 8: ReGEN-TAD performance across prediction horizons. Aggregate detection metrics (AUROC, Average Precision, F1, and Recall) are reported for $H\in\{1,3,5,10,20\}$ . Ranking-based measures (AUROC and AP) remain comparatively stable across horizons, whereas threshold-sensitive metrics (F1 and Recall) exhibit gradual attenuation as forecast uncertainty increases with longer prediction horizons.

DGP	AlioghliOkay	TimeGPT	DAGMM	TGAN-AD	GARCH	IForest	ReGEN-TAD	TranAD	DeepANT	OLS	RRR
Static Factor Model	0.0122	0.0361	0.0607	0.0629	0.0545	0.0921	0.0540	0.0828	0.0429	0.0989	0.1088
Factor-GARCH	0.0056	0.0366	0.0596	0.0544	0.0558	0.0530	0.0538	0.0683	0.0379	0.1039	0.1010
GARCH	0.0000	0.0361	0.0272	0.0418	0.0509	0.0073	0.0528	0.0186	0.0574	0.0977	0.1028
IID Gaussian	0.0037	0.0382	0.0485	0.0579	0.0530	0.0781	0.0526	0.0684	0.0297	0.0877	0.0894
IID Student-t	0.0168	0.0349	0.0636	0.0604	0.0515	0.0597	0.0577	0.0677	0.0286	0.0914	0.0974
Volatility Drift	0.0000	0.0503	0.0000	0.0001	0.0418	0.0000	0.0527	0.0000	0.4279	0.2638	0.2774

Table 6: Average False Positive Rate (FPR) on Clean Data by Data Generating Process (DGP). Models Sorted by FPR (Low to High).

Sector	Anomaly Type	Avg. Match	Sector Size
Information Technology	Volatility spike	0.693	38
Communication Services	Mean shift	0.574	10
Consumer Discretionary	Trend	0.551	18
Health Care	Jump bursts	0.110	12

Table 7: Sector-level attribution recovery under structured perturbations. Anomalies are injected exclusively within a target sector, and recovery is measured via the match-ratio metric, defined as the proportion of affected assets identified among the top-ranked attribution scores.

Figure 9: ReGEN-TAD anomaly score and attribution analysis during two historical systemic crises. Left: 2008 financial crisis. Right: COVID-19 market collapse. In both episodes, anomaly scores remain stable during pre-crisis periods and exhibit sharp concentration during systemic stress intervals. Attribution panels highlight economically coherent cross-sectional contributors consistent with broad market contagion.

Explore similar work

Detecting Time Series Anomalies Like an Expert: A Multi-Agent LLM Framework with Specialized Analyzers
May 7, 2026Hyeongwon Kang, Jeongseob Kim, Jinwoo Park +1Interpretable Anomaly Detection Multi-Agent Systems
Department of Industrial and Management Engineering Korea University · Mirae Asset Securities · Department of Industrial Engineering Seoul National University
Selective Denoising Diffusion Model for Time Series Anomaly Detection
Feb 27, 2026Kohei Obata, Zheng Chen, Yasuko Matsubara +2Time-Series Anomaly Detection Reconstruction-Based Anomaly Detection
SANKEN, The University of Osaka, Osaka, Japan · Great Bay University, Dongguan, China
Temporal-Conditioned Normalizing Flows for Multivariate Time Series Anomaly Detection
Mar 10, 2026David Baumgartner, Helge Langseth, Kenth Engø-Monsen +1Time-Series Anomaly Detection Normalizing Flows
Department of Computer Science, Norwegian University of Science and Technology.

May 7, 2026J/K move · Enter open · S save

Detecting Time Series Anomalies Like an Expert: A Multi-Agent LLM Framework with Specialized Analyzers

Hyeongwon Kang, Jeongseob Kim, Jinwoo Park, Pilsung Kang

Department of Industrial and Management Engineering Korea University · Mirae Asset Securities · Department of Industrial Engineering Seoul National University

Interpretable Anomaly Detection Multi-Agent Systems Time-Series Anomaly Detection

Open paper AskPDF arXiv

Recent studies have explored large language models for time-series anomaly detection, yet existing approaches often rely on a single general-purpose model to directly infer anomaly indices or intervals, limiting controllability, interpretability, and reliability for complex anomaly patterns. We propose SAGE (Specialized Analyzer Group for Expert-like Detection), a multi-agent framework for structured anomaly diagnosis in univariate time series. It decomposes anomaly analysis into four specialized Analyzers for point, structural, seasonal, and pattern anomalies. Each Analyzer applies family-specific numerical tools and diagnostic visualizations to generate evidence, while an evidence-grounded Detector consolidates the evidence into confidence-scored anomaly records with intervals and candidate types. A Supervisor then converts these structured records into analyst-facing diagnostic reports. SAGE further constructs synthetic in-context examples from normal-reference training segments, without using real anomalous segments or anomaly-type labels as in-context examples. Across three benchmarks, SAGE achieves the best average performance among strong ML/DL and language-model-based baselines. Ablation studies and human evaluation further show that the proposed framework improves detection reliability and the practical usefulness of diagnostic outputs.

Abstract

Explore similar work

Detecting Time Series Anomalies Like an Expert: A Multi-Agent LLM Framework with Specialized Analyzers

Selective Denoising Diffusion Model for Time Series Anomaly Detection

Temporal-Conditioned Normalizing Flows for Multivariate Time Series Anomaly Detection