🤖 machine learning

Extreme Volatility Warning under Label Scarcity via Multi-Source Anomaly Fusion

This paper proposes AAMSF, a semisupervised multi-source anomaly fusion framework that outperforms supervised and unsupervised baselines in predicting extreme CSI 300 volatility under severe label scarcity by prioritizing robust anomaly geometry and reliable signal sources over complex supervised representation learning.

Original authors: Jin Qian, Zhangzhi Xiong, Mingrui Li, Zhen Liu

Published 2026-07-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jin Qian, Zhangzhi Xiong, Mingrui Li, Zhen Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the stock market as a giant, chaotic ocean. Most days, the waves are gentle, and the water is calm. But sometimes, without much warning, a massive tsunami hits, wiping out everything in its path. For investors, predicting these "tsunamis"—sudden, extreme crashes or spikes—is the holy grail of risk management. To do this, scientists use "financial data mining," which is basically like trying to find a needle in a haystack by sifting through mountains of information. They look at two main things: the water itself (market prices, trading volume) and the weather reports (news, geopolitical events, social media chatter). The big challenge is that these disasters are incredibly rare. It's like trying to teach a dog to bark only when a specific, once-in-a-lifetime earthquake happens, but you've only ever seen the earthquake happen 80 times in your entire life. If you try to teach the dog with a super-complex, high-tech brain, it gets confused and starts barking at everything. This paper tackles that exact problem: how do you build a warning system for rare financial disasters when you have almost no examples to learn from?

The authors of this paper, Jin Qian and his team, decided to test a common idea: that bigger, smarter computer models are always better. They built a fancy, deep-learning model called HTSF (Hierarchical Text-Signal Fusion) designed to read thousands of news articles and market numbers simultaneously. They hoped this "super-brain" would spot the signs of a coming crash. But here's the twist: the super-brain failed. Because there were so few examples of actual crashes (only about 80 in their training data), the complex model got confused. It started memorizing the noise instead of learning the signal, performing worse than much simpler models. It was like trying to solve a simple math problem using a supercomputer that got stuck in its own gears.

Realizing that "more complexity" was actually the enemy, the team switched tactics. They proposed a new, much simpler system called AAMSF (Anomaly-Augmented Multi-Signal Fusion). Instead of trying to learn a complex language from scratch, this system acts like a vigilant security guard who knows what "normal" looks like. It uses a tool called an "Isolation Forest" (think of it as a machine that spots weird outliers by seeing how hard it is to isolate them from the crowd) to check different sources of information: market numbers, global event data (GDELT), Chinese financial news, and English news.

The team found something surprising about their sources. The Chinese financial news and the global event data were like two best friends who complemented each other perfectly; when one saw trouble, the other usually did too. However, the English news was like a noisy, unreliable friend who kept shouting "Fire!" when there was no smoke. In fact, including the English news actually made the system worse, confusing the guard with false alarms. So, the team built a "lightweight" version that ignored the English news and combined the signals from the market, the Chinese news, and the global events using a simple, fixed rule rather than a complex learning algorithm.

They also added a "time machine" feature called T-AAMSF. They noticed that before a big crash, the warning signs often don't just appear on one day; they build up over two or three days, like a storm getting darker and darker. By letting their system look back at the last few days and add up the "weirdness" scores, they got an even better warning signal.

The results were impressive. On a test set of real market data from 2018 to 2023, their simple, "dumb" system (AAMSF) correctly identified extreme volatility events with a score of 0.680, beating the complex deep-learning models (which scored around 0.588) and even the standard statistical models. Their time-aware version (T-AAMSF) was even better at spotting the specific, rare events, achieving a score of 0.291, which was the highest in their study.

The main lesson from this paper is a bit counter-intuitive: when you are looking for very rare events and have very little data, you don't need a super-complex brain. You need a reliable, simple system that trusts the right sources and ignores the noisy ones. The authors suggest that in the world of financial risk, knowing where to look and how to spot the weirdness is often more important than having a massive, complicated model. They proved that sometimes, the best way to predict a tsunami is not to build a giant, complex wave machine, but to have a simple, sharp eye that knows the difference between a ripple and a real wave.

Technical Summary: Extreme Volatility Warning via Multi-Source Anomaly Fusion

Problem Statement
The paper addresses the challenge of early warning for extreme market volatility in the Chinese capital market (specifically the CSI 300 index), characterized by severe label scarcity. In the studied dataset (2018–2023), there are only approximately 80 positive samples (extreme volatility events) across 791 training days. The task involves predicting whether an extreme volatility event will occur within a 3-day horizon ( $[t+1, t+3]$ ) using a combination of market features and heterogeneous exogenous signals: GDELT geopolitical events, Chinese financial news, and English media coverage. The core difficulty lies in the "rare-event" regime where standard supervised learning approaches struggle due to the extreme mismatch between model complexity and the number of positive labels.

Methodology
The authors propose AAMSF (Anomaly-Augmented Multi-Signal Fusion), a semi-supervised framework designed to minimize learnable parameters, alongside its temporal extension T-AAMSF.

Failure Analysis of Supervised Models: The authors first analyze a 100K-parameter hierarchical text-signal fusion model (HTSF). They demonstrate that increasing architectural complexity (e.g., adding source-level LSTMs and cross-source attention) degrades performance under severe label scarcity. The parameter-to-positive ratio in deep models (approx. 1250:1) leads to overfitting, where additional parameters capture noise rather than signal.
AAMSF Architecture:
- Stage 1 (Unsupervised Anomaly Detection): Instead of learning representations, the model applies Isolation Forest (IF) ensembles independently to each data source (Market, GDELT, Chinese News, English News). Scores are generated using 10-seed ensembles with 200 trees.
- Stage 2 (Lightweight Supervised Component): A Ridge regressor with only 16 parameters is trained on market features to predict the continuous 3-day absolute return ($rv3$). This utilizes all 791 days for training, not just the 80 positives.
- Stage 3 (Score Fusion): The final score is a fixed-weight linear combination: $0.85 \times \text{normalized}(\text{IF}_{\text{combined}}) + 0.15 \times \text{normalized}(\text{Ridge})$ . The IF scores are combined using fixed equal weights ( $1/N$ ).
- Preprocessing: Text features undergo "IF-aware" transformations (rolling z-scores, rate-of-change) to convert absolute values into deviations from recent baselines, making them geometrically suitable for anomaly detection.
T-AAMSF (Temporal Extension): This variant modifies Stage 1 by aggregating IF scores over a multi-day window ( $L$ days) using exponential decay weights ( $w=[0.6, 0.3, 0.1]$ ). This captures the signal build-up typical of extreme volatility events, which often intensify over 2–3 days prior to the event.

Key Findings and Results

Performance: On the held-out test set (2022–2023), AAMSF-Opt (using the curated 3-source set: Market + GDELT + Chinese News) achieves a test AUC-ROC of 0.680. This outperforms the strongest unsupervised baseline (Isolation Forest: 0.630) and the best neural baseline (Simplified TFT: 0.588).
Temporal Gain: T-AAMSF improves the PR-AUC to 0.291, the highest in the study, demonstrating that multi-day anomaly accumulation provides signal beyond same-day snapshots.
Source Asymmetry:
- GDELT and Chinese News: These sources provide complementary risk signals. GDELT is the strongest standalone source (AUC 0.624).
- English News: Contrary to intuition, English media consistently reduces performance (AUC drops to 0.613 when included). Its anomaly structure is anti-correlated with CSI 300 extremes in this setting.
- Weighting Robustness: Learned weighting strategies (e.g., Attention, Stacking) fail because validation data is too noisy (only 242 days) to reliably estimate source quality. For instance, GDELT appears weak on validation (AUC 0.423) but strong on test (AUC 0.625). Consequently, fixed equal weighting outperforms all learned weighting strategies.
Feature Importance: The Ridge component reveals that realized volatility and technical indicators (e.g., Bollinger Band width, intraday range) are the strongest supervised proxies for forward extreme events.

Significance and Claims
The paper argues for a shift in design philosophy for financial risk warning under label scarcity. The primary contribution is the empirical demonstration that robust anomaly geometry and source reliability are more critical than supervised representation capacity in rare-event regimes.

Design Principle: The authors propose that in low-sample regimes, minimizing learnable parameters is essential. They show that a near-parameter-free approach (combining non-parametric anomaly detection with a minimal Ridge regressor) outperforms complex deep learning models.
Practical Insight: The study highlights that "source reliability" is context-dependent; in the Chinese market, domestic and geopolitical event data (GDELT) are superior to global English media for volatility prediction. Furthermore, the paper empirically validates the principle that when detector quality cannot be reliably estimated due to noise, equal weighting is superior to learned weighting.
Limitations: The authors acknowledge that the model is evaluated strictly on the CSI 300 and that cross-market transferability (e.g., to S&P 500 or Nikkei 225) remains untested. They also note that the current evaluation relies on historical splits rather than transaction-cost-aware backtesting.

Technical Summary: Extreme Volatility Warning via Multi-Source Anomaly Fusion

More like this