Introduction
Financial market prediction is a high-stakes challenge characterized by an exceptionally low signal-to-noise ratio. Most price movements resemble random walks, leading standard machine learning models to overfit by attempting to find patterns in pure noise. This paper introduces LARA (Locality-Aware Attention and Iterative Refinement Labeling), a framework designed to identify specific "opportunity" windows where trends are predictable and profitable, rather than attempting to forecast every market tick.
The core philosophy is "Trade when opportunity comes." This approach addresses two primary technical hurdles:
- Selection Bias: Identifying predictable samples without look-ahead bias.
- Label Noise: Managing the stochastic nature of financial labels where identical setups can yield divergent outcomes due to external shocks.
Methodology: The LARA Framework
LARA consists of two primary modules designed to filter noise and refine the learning signal.
Locality-Aware Attention (LA-Attention)
The LA-Attention module serves as a filter to extract potentially profitable samples. It employs a Localization mechanism using masked attention to analyze a sample's neighborhood in the feature space. If a sample is surrounded by historical instances that yielded positive returns, it is flagged as a high-probability opportunity. To optimize this search, Metric Learning is used to define a custom Mahalanobis distance, ensuring that "similarity" is based on features relevant to price movement rather than arbitrary mathematical distance.
Iterative Refinement Labeling (RA-Labeling)
Recognizing that financial labels are often noisy, the RA-Labeling module avoids treating initial labels as absolute truth. It performs Label Refurbishment by iteratively updating labels based on a convex combination of the original label and the model's own predictions. Furthermore, Confidence Weighting utilizes training loss to down-weight samples with high loss, preventing the model from being misled by outliers or "freak accidents" in the data that do not represent repeatable trends.
Experimental Results
The framework was evaluated across three distinct financial domains: China’s A-share stocks (CSI300), high-frequency BTC/USDT cryptocurrency data, and major Chinese ETFs. LARA consistently outperformed state-of-the-art baselines in both precision and profitability.
- Precision: LARA achieved 59.1% precision in stock markets and 57.8% in crypto, leading the best baselines (such as SEAL) by 1.5% to 2.5%.
- Profitability: Back-testing on the 512480.SH ETF demonstrated that LARA's cumulative return significantly exceeds standard models like OLS or LightGBM.
- Ablation Studies: Removing either LA-Attention or RA-Labeling resulted in a marked performance decline, confirming that both opportunity identification and data cleaning are critical to the model's success.
Conclusion
LARA demonstrates that selective prediction—focusing on high-confidence "neighborhoods" in feature space—is more effective than constant prediction in noisy financial environments. While the model is computationally more intensive due to its iterative nature and neighborhood searches, its ability to ignore market chaos and focus on repeatable trends offers a significant advantage for real-world trading strategies.
Summary updated