Explainable Heterogeneous Anomaly Detection in Financial Networks via Adaptive Expert Routing

arXiv:2510.17088 · cs.LG, cs.AI, cs.CE · Submitted 2026-08-16 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Explainable Heterogeneous Anomaly Detection in Financial Networks via Adaptive Expert Routing".

Jane: The paper was written by Zan Li and Rui Fan from Rensselaer Polytechnic Institute.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Title: Tom: Welcome back, everyone. I'm Tom, and with me is Jane. We've got a fascinating paper on the arXiv today, and it's got a bit of a mouthful for a title: "Explainable Heterogeneous Anomaly Detection in Financial Networks via Adaptive Expert Routing."

Jane: Tom, I'll be honest, when I first read that title I had to break it down piece by piece. But once you do, it's actually pretty intuitive. It's about spotting when something goes wrong in the stock market, but not just *that* something went wrong — *why* it went wrong.

Tom: Exactly. And that's the part that got me excited. Most anomaly detection systems just give you a red flag. This one tries to tell you what kind of fire it is before you call the fire department.

Jane: Right. So they're looking at financial networks — think of stocks as nodes in a web, connected by things like sector, geography, or how they move together. When a crisis hits, that web changes shape. This paper builds a system that watches those changes and figures out which of four specific mechanisms is driving the problem.

Tom: And those four mechanisms are straight out of financial economics textbooks. You've got Price-Shock, which is when prices move violently because new information hits. Liquidity, which is when trading just seizes up. Systemic-Contagion, where problems spread from one stock to others like a cold. And Momentum-Reversal, which is when a trend suddenly flips.

Jane: What I love is that they didn't just invent these categories. They grounded them in decades of research. So when the system says "this looks like a liquidity freeze," that's not a random label — it's a diagnosis that points to a specific intervention, like providing market-making support.

Tom: And that's the big deal. A uniform score of zero point nine five could mean two completely different things for two different stocks. One might have a price shock that needs a circuit breaker, the other might have a liquidity problem that needs someone to step in and buy. Same number, totally different response.

Jane: So the title is really promising three things: it's explainable, because it tells you the mechanism. It's heterogeneous, because it treats different failure modes differently. And it uses adaptive routing, which we'll get into later, but it's basically a smart way to decide which specialist should look at each stock.

Tom: And the results are pretty striking. They tested this on one hundred U.S. equities from two thousand seventeen to two thousand twenty-four and it caught all six major stress events in the test period with an average lead time of almost four days. That's not just academic — that's actionable.

Jane: We're going to dig into how they actually built this thing next, because the architecture is clever. But first, I want to flag one thing: the authors are from Rensselaer Polytechnic Institute, and they've clearly thought hard about making this usable in the real world, not just in a lab.

Tom: Stay with us. In the next segment, we'll break down the core problem they're solving and why existing methods just don't cut it for financial networks.

Summary: Tom: So Jane, we've got the title unpacked. Now let's talk about what the paper actually does. The summary in the abstract is dense, but the core idea is that financial anomalies come in different flavors, and you can't treat them all the same way.

Jane: Right. And the authors point out three big challenges that existing methods stumble on. First, financial networks aren't static. When a crisis hits, correlations between stocks change dramatically. A calm market looks very different from a panicked one, and most models use a fixed graph structure that can't adapt.

Tom: That's the adaptivity problem. They show this with a concrete example: during the Silicon Valley Bank collapse, intra-cluster correlations jumped from zero point three one to zero point eight zero. If your model assumes the graph doesn't change, you're flying blind.

Jane: The second challenge is what they call heterogeneity. Different mechanisms — price shocks, liquidity freezes, contagion — produce different statistical signatures. A price shock shows up as fat-tailed returns with stable spreads. A liquidity problem shows up as a bid-ask spread explosion with prices barely moving. Uniform detectors just mush all of that into one scalar score.

Tom: And that's where the "adaptive expert routing" comes in. Instead of one big model trying to do everything, they built four specialized experts, each one tuned to a specific mechanism. Then a routing mechanism decides which expert or combination of experts should handle each stock at each moment.

Jane: The third challenge is interpretability. Most anomaly detectors are black boxes — you get a score, but no explanation. The authors argue that post-hoc explanations, like SHAP values, are unstable and don't tell you which *mechanism* is failing. So they built interpretability directly into the architecture.

Tom: How? The routing weights themselves become the explanation. If the routing weight for the Price-Shock expert spikes, that's the model telling you "this looks like a price shock." You don't need a separate explanation step — the mechanism attribution is part of the forward pass.

Jane: And they validate this with real crises. The SVB collapse in March two thousand twenty-three shows up as a highly localized banking-sector shock, with a forty-four-to-one ratio of routing weight changes in banking versus non-banking stocks. The Japan carry-trade unwind in August two thousand twenty-four on the other hand, shows near-symmetric activation across sectors — a systemic event.

Tom: That distinction is huge. A localized crisis and a systemic crisis call for completely different responses. If you're a regulator, you need to know which one you're dealing with.

Jane: And they wrap it all up in a Market Pressure Index, which aggregates individual stock scores into a market-wide alert with four levels. It's a clean way to go from "this one stock looks weird" to "the whole market is under stress."

Tom: The numbers back it up. They detect all six test events with a three point seven-day average lead time, and they beat the strongest baselines by thirty-three percentage points in detection rate. AUC of zero point eight eight eight, AP of zero point six two six.

Jane: Now, I want to bring in Lu and Meng to get their takes, because this architecture has some interesting trade-offs.

Lu: I'm really taken with the routing weights as interpretability. It's a clever move — you get mechanism attribution for free, without any post-hoc analysis. But I'm curious about the stability of those weights across different market regimes. The paper shows the anomaly score distributions are stable, but what about the routing weights themselves?

Meng: And I want to know about the computational cost. They mention fifty milliseconds per timestep on an A100, which is fine. But the training involves a mixture-of-experts with adaptive temperature and diversity regularization — that's a lot of moving parts. How sensitive is it to hyperparameters?

Jane: Great questions. Let's dig into the methodology next, because the paper actually addresses both of those concerns with ablation studies and sensitivity analysis.

Improvements: Tom: So we've covered the what and the why. Now let's talk about the how — specifically, what this paper improves over existing approaches. Jane, you want to take this?

Jane: Sure. The paper positions itself against three families of methods: temporal models like LSTM-AE and TranAD, static graph methods like DOMINANT, and dynamic graph methods like EvolveGCN and ROLAND. Each has a specific weakness.

Tom: Temporal models treat each stock independently. They miss the contagion effects — the way problems spread through the network. Static graph methods use a fixed adjacency matrix, so they can't adapt when correlations shift. And dynamic graph methods either destabilize under distribution shifts or enforce rigid topologies that miss emergent pathways.

Lu: The key improvement is the stress-modulated adaptive graph fusion. They don't just pick between a prior graph and a learned graph — they blend them with a coefficient that depends on market stress. High stress, you lean on the structural prior. Low stress, you let the data speak.

Meng: That's clever, but it's also a potential failure point. If the stress measure is miscalibrated, the fusion coefficient could be wrong. How do they handle that?

Jane: They clamp the fusion coefficient between zero point two and zero point eight, so neither source ever fully dominates. And they validate the stress measure against four indicators: volatility, pairwise correlation, extreme return fraction, and liquidity intensity. It's not just one noisy signal.

Tom: The second improvement is the mechanism-aligned mixture-of-experts. Instead of one model processing all twenty-nine features uniformly, they partition features into four subsets, one per mechanism. Each expert only sees its own features. That forces specialization.

Meng: And the routing is stress-modulated too. The temperature of the softmax gating increases with stress, so under crisis conditions, the model spreads attention across multiple mechanisms. That makes sense — crises often involve more than one mechanism at once.

Lu: The diversity regularization is also worth noting. They have an entropy floor, a collapse penalty, and an error diversity term. That prevents all experts from converging to the same behavior, which is a classic failure mode in mixture-of-experts.

Jane: And the results show the improvements matter. In the ablation study, removing the stress-modulated fusion drops AUC from zero point eight seven one to zero point seven nine one. Removing the expert specialization entirely — going to a single expert — drops it to zero point seven eight four. Both components contribute independently.

Tom: But the most striking improvement, to me, is the interpretability. The routing weights don't just improve detection — they provide a mechanism attribution that's validated against real crises. The SVB case shows a forty-four-to-one sectoral confinement ratio. The Japan case shows near-symmetric activation. That's not something any baseline can do.

Meng: I'm impressed by the sensitivity analysis too. They show AUC stays above zero point eight five across a wide range of hyperparameters — fusion bounds, expert latent dimensions, entropy coefficients. That suggests the architecture is robust, not just tuned to one configuration.

Lu: And the distribution consistency across regimes is remarkable. The anomaly score distributions are nearly identical across training, validation, and test periods, even though the test period includes a banking crisis and a carry-trade unwind. That's a strong signal that the model learned general mechanisms, not just memorized patterns.

Jane: So the improvements are real and measurable. But I think the biggest contribution is the shift in mindset — from "is there an anomaly?" to "what kind of anomaly is it, and what should we do about it?"

Tom: That's the hook for our final segment. We'll wrap up with what this means for the broader world — regulators, risk managers, and the future of financial AI.

Conclusion: Tom: Alright, let's bring it home. We've been talking about "Explainable Heterogeneous Anomaly Detection in Financial Networks via Adaptive Expert Routing," and I think we've only scratched the surface of why this matters.

Jane: The paper's core contribution is that it doesn't just detect anomalies — it attributes them to specific mechanisms, and it does so without any labeled supervision. That's a big deal for a field where ground truth is scarce and expensive.

Tom: And the empirical results back it up. Six out of six test events detected, with a three point seven-day average lead time. The SVB collapse is identified on the day it happens, which is actually correct — it was an abrupt, information-driven shock. The Japan carry-trade unwind gets a four-day advance warning because it built up gradually.

Lu: I think the theoretical contribution is underappreciated. The routing weights, learned purely from reconstruction error, recover the causal sequence that financial economists have documented for decades: price shocks precede contagion, and liquidity deterioration is a downstream effect, not a leading indicator. That's unsupervised corroboration of crisis transmission theory.

Meng: From a practical standpoint, the fixed detection thresholds are huge. The P95 threshold varies by less than three percent across pandemic, inflation shock, banking stress, and carry-trade unwind. That means you can deploy this without recalibration, which matters for regulatory approval.

Jane: And the implications for practice are concrete. For regulators, it's a tool that can distinguish a localized banking crisis from systemic propagation — that changes the response. For asset managers, the one-to-two-week early warning on the Japan case provides time to reposition, hedge, and pre-position liquidity.

Tom: There's also a broader cultural angle. We're moving toward a world where AI systems don't just flag problems — they explain them in terms that humans can act on. This paper is a step in that direction, and it's grounded in real economic theory, not just pattern matching.

Lalam: I see this as a blueprint for trustworthy AI in high-stakes domains. The architecture embeds interpretability rather than bolting it on afterward. That's the kind of design philosophy we need for any system that touches people's money, health, or safety.

Jane: That's a great point, Lalam. And it's worth remembering that the authors made the code available upon acceptance, which will let other researchers build on this.

Tom: So, to sum up: this paper gives us a mechanism-aware anomaly detection framework that's more accurate, more interpretable, and more stable than anything that came before. It's a genuine advance for financial risk monitoring.

Jane: And with that, we'll say goodbye to this paper. Thanks for listening, and join us next time for another deep dive into the latest research on arXiv.

Tom: Take care, everyone.

Zan Li, Rui Fan

Rensselaer Polytechnic Institute

cs.LG, cs.AI, cs.CE

Submitted: 2026-08-16

Updated: 2026-08-18

Journal ref: XAI-FIN: International Joint Workshop on Explainable AI in Finance, ACM ICAIF 2025

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 56/100

Key concepts

Failure Mechanisms
The system categorizes financial anomalies into four specific types based on financial economics research. These include Price-Shock (violent price movement from new information), Liquidity (trading seizing up), Systemic-Contagion (problems spreading across stocks), and Momentum-Reversal (a sudden trend flip). The model's diagnosis points to a specific intervention needed.
Adaptive Expert Routing
This architecture replaces one large model with four specialized 'experts,' each dedicated to one failure mechanism. A routing mechanism determines which expert(s) should analyze a particular stock at any given moment, allowing the the system to handle diverse types of financial stress appropriately.
Mechanism Attribution
Unlike black-box detectors, this system provides interpretability directly into its design. The routing weights serve as the explanation, indicating precisely which mechanism (e.g., Liquidity) is driving a detected anomaly, providing actionable insight for regulators and risk managers.

Terminology

Summary

arXiv: 2510.17088v2 [cs.LG] 8 Mar 2026

The paper addresses the problem that "financial anomalies arise from heterogeneous mechanisms—price shocks, liquidity freezes, contagion cascades, and momentum reversals—yet existing detectors produce uniform anomaly scores without revealing which mechanism is failing or where risks concentrate. The authors identify three key challenges: (1) static graph structures cannot adapt when correlations shift across regimes; (2) uniform detectors overlook heterogeneous anomaly signatures; and (3) black-box scores provide no actionable guidance on which mechanism drives the anomaly."

The proposed solution is an adaptive graph learning framework that embeds interpretability architecturally rather than post hoc. The framework constructs stress-modulated graphs that adaptively interpolate between known sector and geographic relationships and data-driven correlations as market conditions evolve. Anomalies are decomposed via four mechanism-specific experts—Price-Shock, Liquidity, Systemic-Contagion, and Momentum-Reversal—each capturing a distinct anomaly channel documented in the financial economics literature. The resulting routing weights serve as interpretable proxies for mechanism attribution.

On "100 U.S. equities (2017–2024), the framework detects all six major market stress events with a 3.7-day mean lead time, outperforming the strongest baselines by +33 percentage points in detection rate, with AUC 0.888 and AP 0.626. Case studies on the Silicon Valley Bank collapse (March 2023) and Japan carry-trade unwind (August 2024) demonstrate that routing weights automatically distinguish localized sector-specific crises from systemic multi-sector propagation—without labeled supervision."

The paper motivates the work through two crises: Silicon Valley Bank (SVB) collapsed within 48 hours, triggering 42B in single-day withdrawal attempts and contagion to Signature Bank (March 2023), and Japan's carry-trade unwind sparked a 12% TOPIX crash with global spillovers (August 2024). The fundamental problem is stated: Detecting anomalies without identifying underlying mechanisms provides no actionable guidance. The authors illustrate with two stocks having identical anomaly scores (0.95): "one exhibits a bid-ask spread explosion with stable prices—a liquidity freeze demanding market-making intervention; the other shows extreme return kurtosis with stable spreads—a price shock driven by information asymmetry requiring circuit breakers."

The approach is mechanism-aware detection via specialized expert networks with architectural interpretability, grounding anomaly detection in financial economics by identifying four fundamental mechanisms that explain why market failures occur—Price-Shock, Liquidity, Systemic-Contagion, and Momentum-Reversal.

Financial networks exhibit regime-dependent structure: during crises, sectoral and geographic clustering strengthen; during calm periods, hidden linkages emerge. The paper notes that SVB's sparse tranquil correlations transform into dense banking clusters as distress propagates. Static methods (DOMINANT, CoLA) miss such shifts; dynamic methods face a dilemma: data-driven methods (EvolveGCN, DySAT) often destabilize under distribution shifts; topology-constrained methods (ROLAND) cannot capture emergent contagion pathways.

Solution: A stress-modulated adaptive fusion mechanism constructs the final graph as Afused = αt Aprior + (1 − αt)Alearned, where Aprior encodes static domain knowledge (sectoral and geographic clustering) and Alearned is learned from data. The fusion coefficient αt = clamp(sigmoid(αbase + βα ψt), 0.2, 0.8) is modulated by market stress ψt. Under high stress, αt increases to emphasize Aprior; under low stress, αt decreases to allow Alearned to capture emergent linkages. Clamping αt ∈ [0.2, 0.8] ensures Alearned always contributes.

The four mechanisms have distinct signatures:

  • Price-Shock: "Information-driven volatility from asymmetric news arrival, manifesting as fat-tailed returns (kurtosis >10) with stable spreads. Intervention: Circuit breakers."

  • Liquidity: Trading frictions from market-maker withdrawal, manifesting as bid-ask spread explosion with minimal price movement. Intervention: Liquidity provision.

  • Systemic-Contagion: Cross-market propagation through correlation networks, manifesting as coordinated cross-sector distress with pronounced spread widening. Intervention: Coordinated surveillance.

  • Momentum-Reversal: Regime transitions from trend exhaustion or structural breaks, manifesting as gradual factor loading shifts with RSI sign reversals. Intervention: Portfolio recalibration.

Post-hoc methods (e.g., SHAP, attention visualization) face well-documented limitations: attributions vary across consecutive days, and feature-level statements cannot distinguish liquidity stress from contagion.

  1. Unified mechanism-aware framework integrating stress-modulated adaptive graph fusion, four mechanism-specific experts, and architectural interpretability.

  2. Demonstrated detection superiority on real financial crises with 100% detection of six major events, 3.7-day mean lead time, and a hierarchical Market Pressure Index.

  3. Interpretable mechanism attribution without labeled supervision demonstrated through case studies.

The paper positions against temporal models (LSTM-AE, TranAD, OmniAnomaly, DCdetector, UniTS, MEMTO, CAT), graph-based models (DOMINANT, GDN, EvolveGCN, DySAT, DyGFormer, DyG2Vec, ROLAND, MTAD-GAT, CoLA), and mixture-of-experts approaches (Switch Transformer, V-MoE, GraphMoE). The paper notes that graph-based MoE with memory-augmented routers has been proposed for multivariate time series but approaches that partition the input space by sensor-level or data-structural heterogeneity rather than by financially-grounded failure modes do not yield routing weights interpretable as mechanism attributions.

Regarding interpretability, the paper states that systematic evaluation of post-hoc attribution methods in time-series settings demonstrates temporal instability and feature-level granularity that limit diagnostic value.

Table I compares representative methods across five dimensions (Temporal, Graph, Adaptive, Specialized, Interpretable), showing that Our framework is the first to jointly achieve stress-modulated adaptive graphs, mechanism-aligned expert specialization, and architectural interpretability—no existing method achieves all three simultaneously.

"A financial system with N stocks observed over time is represented as a temporal graph G = (V, Et, Aprior), where V is the set of stocks (V = N), Et denotes time-varying correlations, and Aprior ∈ [0,1]N×N encodes stable sectoral and geographic domain knowledge. Each stock has multivariate features Xi,t ∈ RT×F, where T = 20 (approximately one month of daily data) and F = 29."

The 29 features are partitioned into four mechanism-aligned subsets:

  • Price-Shock (6 features): Volatility and return distribution features capturing information-driven price dislocations

  • Liquidity (8 features): Market microstructure features measuring trading frictions and market-maker activity

  • Systemic-Contagion (7 features): Cross-sectional correlation and spillover features capturing network propagation

  • Momentum-Reversal (8 features): Technical indicators capturing regime transitions and trend exhaustion

Three outputs: (1) entity-level anomaly scores si,t ∈ [0,1]; (2) routing weights wi,t ∈ R4 with Σk wi,t(k) = 1 serving as interpretable proxies for model attention to each mechanism; (3) Market Pressure Index MPIt ∈ [0,1] with four hierarchical alert levels (L1–L4).

Shannon entropy quantifies mechanism concentration: H(wi,t) = −Σk wi,t(k) ln wi,t(k), where Hmin = 0 indicates single-mechanism dominance and Hmax = ln 4 ≈ 1.386 indicates uniform distribution.

Proxy labels derived from mechanism-specific indicators at the 95th percentile threshold support hyperparameter validation:

  • Price-Shock: Extreme returns ri,t

  • Liquidity: Amihud illiquidity ILLIQi,t = ri,t/DollarVolumei,t

  • Systemic-Contagion: Market correlation ρi,tmarket

  • Momentum-Reversal: RSI extremes RSIi,t − 50

These are used only for hyperparameter selection (AUC, AP on validation data) and never affect model training or threshold selection.

Temporal branch: Bidirectional Long Short-Term Memory (BiLSTM) processes input Xi,1:T, concatenating forward and backward hidden states (each 128-dim) to form hbi i,t ∈ RT×256. Multi-head self-attention (4 heads) captures long-range dependencies. Mean pooling and final-step output are concatenated and projected to htemp i,t ∈ R128.

Spatial branch: The full feature window Xi,1:T is flattened to RTF and projected to 128-dim. Two GCN layers propagate information over Aprior: hi(l) = ReLU(Ãprior hi(l−1) Wgcn(l)), l ∈ 1,2, where Ãprior = D−1(Aprior + I). A residual connection is applied: hi(l) ← hi(l−1) + 0.5 hi(l).

Cross-modal fusion: Cross-attention fuses temporal and spatial representations (temporal as query, spatial as key/value), with a residual skip.

Outputs: zi,t = FusionMLP([hfused i,t; hspat i,t]) ∈ R128 (initial embedding) and ci,t = ContextMLP(zi,t) ∈ R64 (historical context).

Market stress: ψt = Σk βk Ĩkt where βk are learnable weights (softmax-normalized), Ĩkt ∈ [0,1] is min-max normalized within each batch. Four raw indicators:

  • I1t = stdi,τ(ri,τ)τ∈[t−T+1,t] (return volatility)

  • I2t = avg pairwise correlation over the window

  • I3t = extreme return fraction (returns exceeding 2σ)

  • I4t = mean liquidity feature intensity

Multi-source graph construction: Atlearned = wtemp Attemp + wctx Atctx + wprior Aprior with temporal similarity Attemp[i,j] = (zi,t·zj,t)/(∥zi,t∥∥zj,t∥) and contextual similarity Atctx[i,j] = (ci,t·cj,t)/(∥ci,t∥∥cj,t∥). An edge gating mechanism modulates edge strengths, and top-k sparsification (k = 20) is applied via a differentiable soft-threshold.

Stress-modulated fusion: αt = clamp(sigmoid(αbase + βα ψt), 0.2, 0.8) and Atfused = αt Aprior + (1 − αt)Atlearned. "High stress (ψt → 1) drives αt → 0.8, emphasizing sectoral and geographic clustering that dominates crisis propagation; low stress (ψt → 0) drives αt → 0.2, allowing learned correlations to capture emergent relationships."

Graph attention refinement: An 8-head GAT refines embeddings: zGAT i,t = Σj∈Nit aij Wgat zj,t, with residual connection zfinal i,t = zi,t + 0.5 zGAT i,t ∈ R128.

Mechanism signal extraction: For each mechanism k, compute f̄i,t(k) = (1/dk) Σj=1..dk fi,T(k)[j] as an inductive prior for routing.

Stress-modulated gating: gi,t = MLP[zfinal i,t; ai,t; ψt; f̄i,t] + bdiv ∈ R4 where bdiv = [0, 0.5, 0.3, 0.2] is a fixed bias favoring Liquidity and Contagion experts. Temperature is stress-modulated: τt = clamp(τbase + βstress ψt, 0.5, 3.0). Routing weights: wi,t = Softmax(gi,t/τt). High stress increases τt, producing softer routing when multiple mechanisms are simultaneously active; low stress sharpens routing toward the dominant mechanism.

Training-time exploration: Gaussian noise ε ∼ N(0, 0.5) is added to gating logits and a minimum weight floor of 0.1 is enforced.

Expert architecture: Each expert encodes its feature subset, fuses with global context, and reconstructs: ei,t(k) = FeatureEncoderk(fi,T(k)) ∈ R64, hi,t(k) = FusionMLPk([ui,t(k); ei,t(k)]) ∈ R128, f̂i,t(k) = Decoderk(hi,t(k)) ∈ Rdk. The Systemic-Contagion expert uses ui,t(k) = zGAT i,t (graph-propagated neighbor information); others use ui,t(k) = zfinal i,t.

Per-expert reconstruction error: li,t(k) = (1/dk)∥fi,T(k) − f̂i,t(k)∥22. Mixture error: eMoE i,t = Σk wi,t(k) li,t(k).

Three parallel MLP decoders reconstruct at horizons h ∈ 1, 3, 5 days. Errors combined with fixed weights: erecon i,t = 0.5 ei,t(1) + 0.3 ei,t(3) + 0.2 ei,t(5).

Entity-level anomaly score: si,t = sigmoid(2 · (0.6 eMoE i,t + 0.4 erecon i,t − µe(b))/(σe(b) + ϵ)) where µe(b), σe(b) are batch statistics.

Market Pressure Index: MPIt = 0.30 mt1 + 0.20 mt2 + 0.30 mt3 + 0.20 mt4 with four components:

  • mt1: mean anomaly rate (weight 0.30)

  • mt2: cross-sectional dispersion (weight 0.20)

  • mt3: tail concentration (weight 0.30)

  • mt4: peak intensity (weight 0.20)

The higher weight on m3 reflects the empirical pattern that extreme anomalies in few entities precede widespread contagion.

Hierarchical alerts: L1 (Observation): MPIt ≥ P70; L2 (Attention): MPIt ≥ P85; L3 (Warning): MPIt ≥ P95; L4 (Crisis): MPIt ≥ P99.

The loss combines: L = 0.4 LMoE + 0.3 Lrec + 0.1 Ldiv + λreg Lreg + 0.05 Laux. Ldiv combines an entropy floor penalizing routing entropy below 0.8·Hmax; a collapse penalty bounding w̄k to [0.15, 0.35]; and an error diversity term. Lreg is lightweight L2 regularization on final embeddings (λreg ≈ 5 × 10−5). Laux comprises a graph sparsity penalty, a spatial-temporal balance loss, and a graph prior consistency penalty. No labeled anomalies are required.

Optimization: AdamW with learning rates η = 5 × 10−4 (main) and ηgraph = 2.5 × 10−3 (graph learner). Cosine annealing with Tmax = 50 epochs. Batch size 32, gradient clipping at norm 1.0, early stopping with patience 10.

Inference: 50ms per timestep on NVIDIA A100 (N = 100 stocks). Model size: 1.85M parameters.

Dataset: "100 U.S. equities drawn from the S&P 500 constituents as of January 2017, with balanced sector representation across all 11 GICS sectors. The dataset spans 1,955 trading days. Temporal splits: Train 2017–2021 (1,240 days), Validation 2022 (232 days), Test 2023–2024 (483 days)." Data sourced from WRDS.

Test events: Six major financial stress events: "SVB Collapse (March 10, 2023), Signature Bank Failure (March 13, 2023), Credit Suisse Crisis (March 20, 2023), Tech Earnings Selloff (October 27, 2023), Weak Jobs Report (August 2, 2024), and Japan Carry-Trade Unwind (August 5, 2024)."

Baselines: Temporal (LSTM-AE, TranAD, OmniAnomaly), Static graph (DOMINANT, GDN, AnomalyDAE), Dynamic graph (EvolveGCN, MTAD-GAT, ROLAND), MoE (GraphMoE).

"Our framework detects all six events with a 3.7-day mean lead time, outperforming all baselines by +33 percentage points over the strongest methods (TranAD, GDN, MTAD-GAT, and ROLAND, each detecting 4 of 6 events). Test AUC 0.888 and AP 0.626."

"Notably, the SVB collapse registers a lead time of 0 days—the threshold is crossed on the event date itself rather than in advance. This is consistent with SVB's character as a localized banking-sector shock: the absence of pre-event market-wide stress buildup is precisely what distinguishes it from systemic crises such as the Japan carry-trade unwind (4-day lead), and is itself a meaningful model output rather than a detection failure."

Ablating stress-modulated fusion: "pure Aprior (AUC 0.712) misses emergent correlations; pure Alearned (AUC 0.651) overfits to volatile crisis-time correlations; fixing α = 0.5 without stress modulation (AUC 0.791) isolates the contribution of adaptive rebalancing (−0.080 AUC vs. full model)."

Replacing stress-aware MoE: single expert (AUC 0.784) or uniform routing weights (AUC 0.803) — "the larger gap from removing partitioning entirely (−0.087 AUC, single expert) relative to removing adaptive routing (−0.068 AUC, uniform weights) indicates that mechanism-specific feature partitioning is the more critical component."

SVB collapse (March 10, 2023): Banking stocks show explosive Price-Shock (+88%) and Systemic-Contagion (+44%) activation, while non-banking stocks remain near baseline (Price-Shock +2%, Systemic-Contagion +5%), yielding a confinement ratio of 44:1. MPI spikes sharply at the event date with no prior elevation.

Japan carry-trade unwind (August 5, 2024): Both sectors show substantial and symmetric activation—non-banking exhibits higher Price-Shock (+39% vs. +31%) and Systemic-Contagion (+29% vs. +11%) than banking—with a confinement ratio of ≈1:1. MPI crosses the alert threshold on Aug 1, providing a 4-day advance warning.

Normal-regime weights (< P75): [0.301, 0.415, 0.169, 0.115] shift to [0.614, 0.100, 0.256, 0.030] under anomaly conditions (> P95). "Price-Shock +104%: Volatility features (kurtosis, VaR violations) dominate during crises. Liquidity −76%: Liquidity deterioration is a downstream consequence of price shocks rather than a primary causal signal. Systemic-Contagion +52%: Activates when correlations surge but remains secondary to volatility. Momentum-Reversal −74%: Technical indicators lose predictive power during structural dislocations."

Mean scores differ by ≤0.010 and P95 thresholds vary ≤3% across splits. Summary statistics: mean [0.4435, 0.4340, 0.4383], std [0.2387, 0.2339, 0.2354], P95 [0.928, 0.955, 0.944], P99 [0.989, 0.997, 0.996]. Validation-calibrated thresholds (P95=0.955, P99=0.997) remain stable on the test set (P95=0.944, P99=0.996, ≤1% drift).

"AUC and AP remain stable across wide hyperparameter ranges: fusion bounds αt ∈ [0.15, 0.85] (∆AUC ≤0.003), expert latent de ∈ [32, 128] (∆AUC ≤0.003), entropy coefficient γ1 ∈ [0.005, 0.02] (∆AUC ≤0.005). Performance degrades only at extremes."

Computational efficiency: our method requires 10.2 sec/epoch and 50 ms per inference step on an NVIDIA A100, comparable to leading baselines (TranAD: 8.7 sec/42 ms; ROLAND: 11.5 sec/58 ms; MTAD-GAT: 9.3 sec/47 ms).

The paper presents "a mechanism-aware framework for anomaly detection in dynamic financial networks, resolving three fundamental challenges through principled architectural design grounded in financial economics: the stability-responsiveness dilemma in adaptive graph construction, the attribution gap between anomaly detection and mechanism identification, and the capacity inefficiency of uniform detectors facing heterogeneous failure modes."

Empirical contributions: 100% detection across all six major stress episodes in the test period with a 3.7-day mean lead time, AUC 0.888, and AP 0.626—outperforming the strongest baselines by +33pp in detection rate, +30% in AUC, and +62% in AP. The SVB collapse exhibits a 44:1 sectoral confinement ratio consistent with a localized banking-sector shock, while the Japan carry-trade unwind shows near-symmetric cross-sector activation consistent with systemic propagation.

Implications for financial practice: "The SVB collapse is correctly identified on the event date itself—consistent with its character as an abrupt, information-driven shock rather than a gradually accumulating systemic stress—while the Japan carry-trade unwind triggers a 4-day advance warning." Fixed detection thresholds stable across heterogeneous regimes (P95 variance ≤3%) support consistent, auditable decision criteria.

Contributions to financial economics theory: "Because routing weights emerge from unsupervised reconstruction training with no access to mechanism labels or crisis annotations, their learned ordering constitutes unsupervised empirical corroboration of crisis transmission theory: the model, guided only by reconstruction error, recovers the same causal sequence that financial economists have documented through structural analysis—Price-Shock precedes Systemic-Contagion (volatility triggers network effects), and Liquidity deterioration is a downstream consequence rather than a leading indicator."

Improvements for AI systems

Based on the paper, here are the specific improvements I can make to AI systems and the resulting capabilities:

Improvement: Replace uniform anomaly detectors with a mixture-of-experts architecture where each expert specializes in a distinct failure mechanism (Price-Shock, Liquidity, Systemic-Contagion, Momentum-Reversal), with routing weights serving as interpretable attribution signals.

Resulting capability: The AI system can now distinguish why an anomaly occurs, not just that it occurs. For example, two stocks with identical anomaly scores (0.95) are now correctly differentiated: one flagged as a liquidity freeze (requiring market-making support) versus a price shock (requiring circuit breakers).

Improvement: Implement dynamic graph fusion where the adjacency matrix adapts based on market stress: Afused = αt·Aprior + (1−αt)·Alearned, with αt clamped to [0.2, 0.8] and modulated by a composite stress index.

Improvement: Embed interpretability directly into the model structure rather than applying post-hoc explanation methods (SHAP, attention visualization) that suffer from temporal instability.

Improvement: Aggregate entity-level anomaly scores into a four-component market-level index (mean rate, dispersion, tail concentration, peak intensity) with four alert levels (L1–L4) calibrated to validation-set percentiles.

Improvement: Use routing weight dynamics relative to baseline periods to automatically classify crisis scope without labeled supervision.

Improvement: Design the training objective and architecture so that anomaly score distributions remain stable across heterogeneous market regimes (pandemic, inflation shock, banking stress, carry-trade unwind).

Improvement: Partition 29 financial features into four mechanism-aligned subsets (Price-Shock: 6 features, Liquidity: 8, Systemic-Contagion: 7, Momentum-Reversal: 8), with each expert processing only its mechanism-relevant features.

Improvement: Enforce balanced expert training through an entropy floor, routing collapse penalty, and error diversity term, preventing any single expert from dominating.

  1. Detect all six major market stress events in the 2023–2024 test period with 100% detection rate and 3.7-day mean lead time (vs. 66.7% for best baselines).

  2. Provide 4-day advance warning for systemic crises (Japan carry-trade unwind) and 1–2-week stock-level early signals, enabling pre-positioning of liquidity, hedging, and portfolio restructuring.

  3. Automatically distinguish crisis types without labeled data: localized banking shocks (SVB) versus systemic global contagion (Japan), enabling targeted interventions (circuit breakers vs. coordinated surveillance).

  4. Maintain stable performance across heterogeneous market regimes (AUC 0.888, AP 0.626 on test set) with fixed thresholds, eliminating recalibration costs.

  5. Achieve computational efficiency comparable to simpler baselines (50ms inference, 1.85M parameters), making interpretability and mechanism-awareness computationally affordable.

  6. Provide unsupervised empirical corroboration of crisis transmission theory: the system independently recovers the documented causal sequence (Price-Shock precedes Systemic-Contagion; Liquidity is downstream), validating the architecture's financial grounding.

Related papers