Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Explainable Heterogeneous Anomaly Detection in Financial Networks via Adaptive Expert Routing

Zan Li, Rui Fan

TL;DR

The paper tackles explainable anomaly detection in financial networks by addressing heterogeneity in anomaly mechanisms and regime changes. It introduces an adaptive graph learning framework with stress-modulated fusion and a four-way mixture-of-experts that routes observations to mechanism-specific predictors, enabling dual-level interpretability through static mechanism attribution and dynamic weight trajectories. Empirically, it achieves a 92.3% detection rate across 13 major market events with a 3.8-day lead time, outperforming strong baselines, and demonstrates interpretable mechanism evolution in the Silicon Valley Bank case. The approach has practical implications for regulators and market participants by providing targeted insights into which mechanisms are active and how they evolve over time.

Abstract

Financial anomalies exhibit heterogeneous mechanisms (price shocks, liquidity freezes, contagion cascades, regime shifts), but existing detectors treat all anomalies uniformly, producing scalar scores without revealing which mechanism is failing, where risks concentrate, or how to intervene. This opacity prevents targeted regulatory responses. Three unsolved challenges persist: (1) static graph structures cannot adapt when market correlations shift during regime changes; (2) uniform detection mechanisms miss type-specific signatures across multiple temporal scales while failing to integrate individual behaviors with network contagion; (3) black-box outputs provide no actionable guidance on anomaly mechanisms or their temporal evolution. We address these via adaptive graph learning with specialized expert networks that provide built-in interpretability. Our framework captures multi-scale temporal dependencies through BiLSTM with self-attention, fuses temporal and spatial information via cross-modal attention, learns dynamic graphs through neural multi-source interpolation, adaptively balances learned dynamics with structural priors via stress-modulated fusion, routes anomalies to four mechanism-specific experts, and produces dual-level interpretable attributions. Critically, interpretability is embedded architecturally rather than applied post-hoc. On 100 US equities (2017-2024), we achieve 92.3% detection of 13 major events with 3.8-day lead time, outperforming best baseline by 30.8pp. Silicon Valley Bank case study demonstrates anomaly evolution tracking: Price-Shock expert weight rose to 0.39 (33% above baseline 0.29) during closure, peaking at 0.48 (66% above baseline) one week later, revealing automatic temporal mechanism identification without labeled supervision.

Explainable Heterogeneous Anomaly Detection in Financial Networks via Adaptive Expert Routing

TL;DR

The paper tackles explainable anomaly detection in financial networks by addressing heterogeneity in anomaly mechanisms and regime changes. It introduces an adaptive graph learning framework with stress-modulated fusion and a four-way mixture-of-experts that routes observations to mechanism-specific predictors, enabling dual-level interpretability through static mechanism attribution and dynamic weight trajectories. Empirically, it achieves a 92.3% detection rate across 13 major market events with a 3.8-day lead time, outperforming strong baselines, and demonstrates interpretable mechanism evolution in the Silicon Valley Bank case. The approach has practical implications for regulators and market participants by providing targeted insights into which mechanisms are active and how they evolve over time.

Abstract

Financial anomalies exhibit heterogeneous mechanisms (price shocks, liquidity freezes, contagion cascades, regime shifts), but existing detectors treat all anomalies uniformly, producing scalar scores without revealing which mechanism is failing, where risks concentrate, or how to intervene. This opacity prevents targeted regulatory responses. Three unsolved challenges persist: (1) static graph structures cannot adapt when market correlations shift during regime changes; (2) uniform detection mechanisms miss type-specific signatures across multiple temporal scales while failing to integrate individual behaviors with network contagion; (3) black-box outputs provide no actionable guidance on anomaly mechanisms or their temporal evolution. We address these via adaptive graph learning with specialized expert networks that provide built-in interpretability. Our framework captures multi-scale temporal dependencies through BiLSTM with self-attention, fuses temporal and spatial information via cross-modal attention, learns dynamic graphs through neural multi-source interpolation, adaptively balances learned dynamics with structural priors via stress-modulated fusion, routes anomalies to four mechanism-specific experts, and produces dual-level interpretable attributions. Critically, interpretability is embedded architecturally rather than applied post-hoc. On 100 US equities (2017-2024), we achieve 92.3% detection of 13 major events with 3.8-day lead time, outperforming best baseline by 30.8pp. Silicon Valley Bank case study demonstrates anomaly evolution tracking: Price-Shock expert weight rose to 0.39 (33% above baseline 0.29) during closure, peaking at 0.48 (66% above baseline) one week later, revealing automatic temporal mechanism identification without labeled supervision.
Paper Structure (29 sections, 37 equations, 3 figures, 8 tables)

This paper contains 29 sections, 37 equations, 3 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Method
  4. Overview
  5. Module 1: Spatial-Temporal Encoding
  6. Module 2: Neural Dynamic Graph Learning & Adaptive Fusion
  7. Module 3: Stress-Modulated Mixture-of-Experts
  8. Module 4: Multi-Scale Reconstruction & Market Pressure Index
  9. Experiments
  10. Experimental Setup
  11. Overall Performance
  12. Silicon Valley Bank Case Study
  13. Ablation Study
  14. Conclusion
  15. Complete Mathematical Formulations
  16. ...and 14 more sections

Figures (3)

  • Figure 1: Three fundamental challenges in financial anomaly detection.(a) Challenge 1—Adaptivity: Temporal network evolution during crisis. Left: tranquil period with baseline correlations (avg 0.35); Middle: localized shock at node 3; Right: contagion spreads with correlation spikes (0.37) and new high-correlation clusters (thick red edges). Static graphs treat all phases identically; purely data-driven dynamic graphs destabilize. (b) Challenge 2—Specialization: Four distinct anomaly signatures. From left: distributional extremes (four-sigma fat tails, price shocks), localized microstructure breakdowns (spread spikes, liquidity crises), cascading contagion (progressive infections), prolonged transitions (momentum reversals, regime shifts). Uniform detection cannot distinguish mechanisms. (c) Challenge 3—Interpretability: Scalar scores (top) provide no mechanism attribution. Our MoE architecture (bottom) routes to specialized experts—routing weight concentration on Systemic-Contagion expert (colored boxes) reveals active mechanism, enabling targeted intervention.
  • Figure 2: Four-module framework architecture. Module 1 extracts multi-scale spatial-temporal representations. Module 2 (addressing Challenge 1) learns dynamic graphs with stress-modulated adaptive fusion. Module 3 (addressing Challenge 2) routes observations to four specialized experts with adaptive temperature. Module 4 (addressing Challenge 3) produces interpretable outputs: individual scores $s$, routing weights $w$ for mechanism attribution, and Market Pressure Index with hierarchical alerts.
  • Figure 3: SVB anomaly evolution tracking (March 2023). Baseline values computed from February 8-18 (days -30 to -20 before event) average. (a) Market Pressure Index elevated to 0.60 on March 10 (event day)—58% increase versus baseline 0.38, outperforming individual scores (32% increase). (b) Crisis localization: banking sector stocks ($n=30$) peaked at mean score 0.87 (149% above baseline); non-banking stocks ($n=70$) reached 0.44 (19% above baseline). (c) Expert weight evolution: Price-Shock expert (red) rose from baseline 0.29 to 0.39 upon closure (33% elevation), peaked at 0.48 on March 17 (66% elevation); Systemic-Contagion expert (green) maintained 0.13 (16% above baseline); Liquidity (blue) and Momentum-Reversal (orange) declined. Weight trajectories reveal progression: isolated shock (March 9) $\to$ intensification (March 10-12) $\to$ contagion spread (March 13-17).