Explainable Heterogeneous Anomaly Detection in Financial Networks via Adaptive Expert Routing

Different financial crises require different solutions: The paper identifies four distinct anomaly mechanisms (price shocks, liquidity freezes, contagion cascades, and momentum reversals), demonstrating that a one-size-fits-all anomaly detection system cannot effectively guide regulatory responses. Liquidity freezes need market-making support while information-driven price shocks need circuit breakers—regulators must know which mechanism is occurring to respond appropriately.

Explainability should be built into AI systems from the start, not added later: Rather than using "black-box" models that produce unexplained anomaly scores, this framework embeds interpretability into its core architecture through mechanism-specific experts. This architectural approach to explainability is more reliable than post-hoc explanation methods and directly supports decision-making by regulators and market participants.

AI systems must adapt to changing market conditions: The framework uses "stress-modulated graphs" that adjust their understanding of financial relationships as market conditions evolve, rather than relying on static historical patterns. This addresses a critical gap in existing systems that cannot account for shifting correlations during different market regimes.

Early warning systems can provide actionable lead time: The framework detected all six major market stress events in the 2017-2024 period with a mean lead time of 3.7 days, suggesting that interpretable AI systems could give regulators and market participants meaningful time to implement protective measures before crises fully materialize.