Volatility-Aware Extreme Event Detection in High-Frequency Financial Markets

Predicting extreme price movements in high-frequency financial markets is notoriously difficult due to non-stationarity, microstructure noise, and severe class imbalance. Most existing research focuses on developing increasingly complex model architectures to solve these problems. This study shifts the focus from model complexity to data-centric problem formulation, specifically investigating whether redefining the prediction target can improve the detection of rare, high-impact events in Bitcoin limit order book (LOB) data.

The authors propose a volatility-aware target formulation. Instead of defining extreme events solely based on large future returns—which results in a highly imbalanced dataset where extreme events are rare—the authors incorporate high-volatility regimes into the labeling process. This approach is motivated by the empirical observation of volatility clustering in financial markets. By combining return magnitude with volatility indicators, the researchers increase the proportion of informative samples. They utilize an XGBoost model within a rigorous time-series cross-validation framework to evaluate the performance of this new target definition against a standard baseline.

The study demonstrates that target design is a critical, often overlooked, driver of predictive performance in financial machine learning. The volatility-aware target formulation increased the proportion of positive samples from approximately 2% to 6%, effectively mitigating some of the challenges posed by class imbalance. Consequently, the proposed method achieved a Precision-Recall AUC of approximately 0.40, compared to a PR-AUC of 0.06 for the baseline. This result highlights that how a problem is framed and how data is labeled can have a more substantial impact on model success than the choice of a complex architecture.