股市时间序列异常检测：文献缺失原因及资源问询

Hey, great question—your observation about the scarcity of stock market time-series anomaly detection papers compared to prediction work is spot-on. Let’s unpack why this is the case, and then point you to some useful resources to dive deeper:

Why Are Published Papers So Rare? #

1. Unique Complexities of Stock Market Data #

• Ambiguous anomaly definitions: Unlike earthquakes (clear magnitude thresholds) or IDS (distinct attack signatures), "anomalies" in stocks have no universal standard. Is a 10% single-day drop an anomaly? Or a 3-day deviation from the moving average by 2σ? Different investors, regulators, and firms define anomalies entirely differently, making it hard to create shared benchmark datasets or evaluation metrics—key for academic consensus.
• Noise and dynamicity: Stock prices are driven by policy shifts, public sentiment, and black-swan events. What looks like an anomaly might just be temporary market sentiment, not a true "abnormal" pattern. Plus, market rules and participant behavior evolve constantly, so models struggle to generalize—making academic results less actionable.
• Data barriers: Most high-value anomaly data (e.g., high-frequency trading irregularities, insider trading traces) is held by exchanges, brokerages, or regulators, not publicly available. Without labeled, high-quality data, anomaly detection research (which relies heavily on validation) is tough to pull off.

2. Hidden Value & Commercial Incentives #

• Prediction has more obvious commercial appeal: Everyone from retail traders to hedge funds wants to predict prices to make money. This drives funding and publication, since results can be measured directly by profit metrics.
• Anomaly detection is often risk-focused: Much of this work is done for internal use—regulators monitoring for manipulation, funds checking for portfolio irregularities. Publishing these methods would let bad actors avoid detection, so firms keep this research private instead of submitting it to journals.

Resources to Explore #

• arXiv’s q-fin Category: Don’t limit yourself to Google/Semantic Scholar. The Quantitative Finance section on arXiv has unpublished preprints covering stock anomaly detection—look for work on transformer-based high-frequency anomaly spotting, or volatility-clustering detection methods.
• Quantitative Firm Blogs: Top hedge funds and proprietary trading firms sometimes share risk-focused anomaly detection insights on their tech blogs. While they won’t publish full code, these posts give you a look at real-world, production-grade approaches (often combining statistical methods like EWMA/ARIMA residuals with ML models like Isolation Forest or Autoencoders).
• GitHub Open Source Projects: Search for stock time series anomaly detection on GitHub. You’ll find plenty of practitioner-shared projects using public datasets to test anomaly detection pipelines—great for hands-on learning.
• Specialized Books:
  • Quantitative Risk Management by Alexander J. McNeil, Rudiger Frey, and Paul Embrechts: Has dedicated chapters on outlier detection and risk-focused time-series analysis for finance.
  • Advances in Financial Machine Learning by Marcos López de Prado: Covers practical anomaly handling techniques tailored to financial data, including how to label anomalies for training.

内容的提问来源于stack exchange，提问作者O. Yigit