Extreme Volatility Warning under Label Scarcity via Multi-Source Anomaly Fusion
Proposed AAMSF framework for extreme volatility warning via multi-source anomaly fusion; achieved test AUC of 0.680.
Key Findings
Methodology
AAMSF combines Isolation Forest anomaly detection and lightweight Ridge regression with fixed-weight fusion across market indicators, GDELT events, Chinese financial news, and English media. T-AAMSF extends this by aggregating multi-day anomaly signals.
Key Results
- AAMSF achieved test AUC of 0.680 on CSI 300, outperforming unsupervised baseline (0.630) and neural baseline (0.588).
- T-AAMSF improved PR-AUC to 0.291 via multi-day anomaly accumulation, outperforming single-day methods.
- Ablation studies revealed GDELT and Chinese news signals are complementary, while English media consistently reduced performance.
Significance
Addresses label scarcity in financial extreme event warning by introducing a lightweight anomaly-based framework, suitable for noisy, low-sample scenarios. Impacts both academia and industry by providing robust risk warning mechanisms.
Technical Contribution
Introduced an anomaly-geometry-based multi-source fusion framework, avoiding overfitting in label-scarce scenarios while enhancing robustness through fixed-weight design.
Novelty
First to combine multi-source anomaly detection with lightweight score fusion for financial extreme event warning, extending signal capture through temporal aggregation.
Limitations
- Evaluated only on CSI 300; cross-market generalization remains untested.
- Fixed-weight design may not adapt to dynamic signal changes.
- Does not account for transaction costs or policy shifts affecting warning systems.
Future Work
Future work can explore cross-market generalization, optimize temporal aggregation strategies, and integrate dynamic weighting for improved adaptability.
AI Executive Summary
Extreme market volatility warning is central to financial risk management, yet traditional methods struggle with label scarcity and heterogeneous signals.
This paper introduces AAMSF, a framework combining Isolation Forest anomaly detection with lightweight Ridge regression and fixed-weight fusion across market indicators, GDELT events, Chinese financial news, and English media. Experiments demonstrate AAMSF achieves test AUC of 0.680 on CSI 300, outperforming unsupervised and neural baselines.
Additionally, T-AAMSF improves PR-AUC to 0.291 by aggregating multi-day anomaly signals, validating the effectiveness of temporal accumulation. The study highlights the importance of anomaly geometry and signal reliability, offering new insights for label-scarce financial risk warning systems.
Deep Analysis
Background
Extreme volatility in financial markets is often triggered by exogenous shocks. Traditional price-based models like GARCH and EVT fail to leverage textual and event signals. Recent multi-source fusion methods focus on routine return prediction but overlook label-scarce extreme event forecasting.
Core Problem
In the CSI 300 market, extreme volatility events are rare, with only 80 positive samples. Deep learning models overfit in such scenarios. The challenge lies in achieving reliable warning under label scarcity and signal heterogeneity.
Innovation
AAMSF combines multi-source anomaly detection with lightweight score fusion, avoiding overfitting in deep models. T-AAMSF further enhances signal capture through temporal aggregation over multi-day windows.
Methodology
- �� Use Isolation Forest for anomaly detection across market indicators, GDELT events, Chinese news, and English news.
- �� Employ Ridge regression to predict continuous rv3 using all training days.
- �� Fixed-weight fusion of multi-source signals to mitigate validation noise.
- �� Extend AAMSF with T-AAMSF by aggregating multi-day anomaly signals using exponential decay weights.
Experiments
Experiments use CSI 300 data (2018–2023) with strict temporal splits for training, validation, and testing. Baselines include GARCH, Isolation Forest, and neural models. Metrics include AUC and PR-AUC.
Results
AAMSF achieved test AUC of 0.680, outperforming unsupervised baseline (0.630) and neural baseline (0.588). T-AAMSF improved PR-AUC to 0.291, validating the effectiveness of temporal aggregation strategies.
Applications
Applicable for financial institutions' risk warning systems, enabling real-time monitoring of market volatility and supporting investment decision-making.
Limitations & Outlook
Model has not been tested for cross-market generalization. Fixed-weight design may not adapt to dynamic signal changes. Transaction costs and policy shifts are not accounted for.
Plain Language Accessible to non-experts
Imagine a factory receiving raw materials (market data, news events) daily. Workers (algorithms) must decide if a major accident might occur. AAMSF acts like experienced workers, quickly inspecting material quality (anomaly detection) and applying simple rules (fixed weights) for reliable warnings. T-AAMSF observes trends over multiple days, improving accuracy further.
ELI14 Explained like you're 14
Imagine you're planning a school event, but the weather might turn bad! You check the school's forecast, news reports, and teachers' advice. AAMSF is like a smart assistant that combines all this info to tell you if you should cancel the event. T-AAMSF is even cooler—it looks at trends over the past few days to give an even better recommendation!
Glossary
Isolation Forest
An unsupervised anomaly detection algorithm based on random partitioning path lengths.
Used to detect anomalies in market indicators and news signals.
Ridge Regression
A regularized linear regression method to prevent overfitting.
Used to predict the continuous target rv3.
GDELT
A global event database recording sentiment, conflict, and cooperation metrics.
Provides external event signals.
AUC
Area under the curve, a metric for classifier performance.
Used to compare predictive capabilities of models.
PR-AUC
Precision-recall area under the curve, evaluates classifier performance on imbalanced data.
Measures accuracy of extreme event warnings.
Open Questions Unanswered questions from this research
- 1 How can AAMSF be validated across different markets?
- 2 How can dynamic weighting adapt to signal changes?
Applications
Immediate Applications
Financial Risk Warning
Real-time monitoring of market volatility to help investors avoid risks.
Policy Sensitivity Analysis
Analyze the impact of policy changes on markets for decision support.
Long-term Vision
Cross-Market Risk Monitoring
Develop global risk warning systems to enhance financial stability.
Abstract
Early warning of extreme market volatility is central to financial risk management, but actionable events are rare, nonstationary, and often triggered by exogenous information shocks. In our CSI~300 setting, only $\sim$80 positive samples are observed across 791 training days, making heavily supervised multi-source models unstable. We first analyze a 100K-parameter hierarchical text-signal fusion model (HTSF) and find that added parameterization hurts in this low-label regime. Motivated by this failure, we propose \textbf{AAMSF} (Anomaly-Augmented Multi-Signal Fusion), a semisupervised framework that combines Isolation Forest anomaly scores over market indicators, GDELT events, Chinese financial news, and English media with lightweight Ridge score fusion. We further introduce \textbf{T-AAMSF}, a temporal extension for multi-day anomaly accumulation. On CSI~300 (2018--2023), AAMSF achieves test AUC-ROC \textbf{0.680}, outperforming the strongest unsupervised baseline (0.630) and neural baseline (0.588), while T-AAMSF improves PR-AUC to 0.291. Ablations reveal strong source asymmetry: GDELT and domestic financial news provide complementary risk signals, whereas English media consistently reduces performance, and learned weighting is unreliable under validation noise. These results suggest an empirical design principle for label-scarce financial risk warning: robust anomaly geometry and source reliability can matter more than supervised representation capacity.