Abstract
Price movement forecasting aims at predicting the future trends of financial assets based on the current market conditions and other relevant information. Recently, machine learning(ML) methods have become increasingly popular and achieved promising results for price movement forecasting in both academia and industry. Most existing ML solutions formulate the forecasting problem as a classification(to predict the direction) or a regression(to predict the return) problem over the entire set of training data. However, due to the extremely low signal-to-noise ratio and stochastic nature of financial data, good trading opportunities are extremely scarce. As a result, without careful selection of potentially profitable samples, such ML methods are prone to capture the patterns of noises instead of real signals. To address this issue, we propose a novel price movement forecasting framework, called Locality-Aware Attention and Iterative Refinement Labeling(LARA), which consists of two main components: 1)Locality-aware attention automatically extracts the potentially profitable samples by attending to surrounding class-aware label information. Moreover, equipped with metric learning techniques, locality-aware attention enjoys task-specific distance metrics and distributes attention on potentially profitable samples in a more effective way. 2)Iterative refinement labeling further iteratively refines the labels of noisy samples and then combines the learned predictors to be robust to the unseen and noisy samples. In a number of experiments on three real-world financial markets: ETFs, stocks, and cryptocurrencies, LARA achieves superior performance compared with the traditional time-series analysis methods and a set of machine learning based competitors on the Qlib platform. Extensive ablation studies and experiments also demonstrate that LARA indeed captures more reliable trading opportunities.
Introduction
Predicting financial market movements is a high-stakes challenge characterized by an exceptionally low signal-to-noise ratio. Standard machine learning approaches typically attempt to learn from every data point in a time series. However, financial markets often exhibit stochastic behavior akin to random walks, where clear, profitable signals are the exception rather than the rule. When models are forced to learn from these "noisy" periods, they frequently overfit, mistaking random fluctuations for predictive patterns.
The LARA framework (Locality-Aware Attention and Iterative Refinement Labeling) addresses this by shifting the focus from global learning to the identification of rare, high-probability trading opportunities. It tackles two fundamental challenges: effectively extracting these "opportunities" from noisy datasets and adaptively refining labels that may be corrupted by freak market events.
The LARA Framework
LARA introduces a two-stage pipeline designed to filter noise and enhance signal robustness through selective sample extraction and label refinement.
Locality-Aware Attention (LA-Attention)
The LA-Attention module serves as an intelligent filter. It employs Metric Learning to derive a specialized distance metric (Mahalanobis distance) that clusters similar market states while separating dissimilar ones. A subsequent Localization Module utilizes a masked attention mechanism to estimate the profitability of a sample based on its "neighbors" in the learned embedding space. Samples exceeding a specific probability threshold are flagged as actionable opportunities, effectively ignoring the surrounding market noise.
Iterative Refinement Labeling (RA-Labeling)
Recognizing that even identified opportunities may have "noisy" labels (e.g., a strong signal resulting in a loss due to exogenous shocks), LARA implements an iterative "refurbishment" process. The framework adaptively adjusts labels based on a convex combination of the original label and the model's high-confidence predictions. This iterative process generates an ensemble of predictors that are significantly more robust to market volatility and label noise.
Experimental Evaluation
LARA was evaluated across three diverse financial markets: China’s A-share stocks (CSI300 index constituents), high-frequency cryptocurrency data (BTC/USDT from OKEx), and major Chinese ETFs.
The framework was compared against a broad spectrum of baselines, including specialized quantitative models (TabNet, Adamct, LightGBM), state-of-the-art time-series models (iTransformer, PatchTST), and noisy-label methods (SEAL, FINE).
Key Results
- Superior Precision: On the CSI300 dataset, LARA achieved a precision (PR) of 59.1%, a 2.5% improvement over the strongest baseline (SEAL). In the cryptocurrency market, it reached 57.8% PR.
- Enhanced Returns: LARA consistently delivered higher Average Returns (AR) across all tested asset classes.
- Ablation Insights: Studies confirmed that applying LA-Attention during both training and testing is vital; removing it from either phase resulted in precision drops of up to 40%.
Conclusion
The LARA framework demonstrates that in financial forecasting, "less is more." By selectively trading only when high-confidence opportunities arise and iteratively refining training labels, LARA overcomes the noise-trap that hinders traditional global-learning models. Future work will focus on optimizing hyperparameter sensitivity and extending the framework to continuous return regression.
Summary updated