Trade When Opportunity Comes: Price Movement Forecasting via Locality-Aware Attention and Iterative Refinement Labeling

TL;DR

The paper tackles price movement forecasting under extreme noise by introducing LARA, a two-component framework that first uses Locality-Aware Attention (LA-Attention) to selectively extract potentially profitable samples via a learned metric and local neighborhood attention, and then applies Iterative Refinement Labeling (RA-Labeling) to robustly refine noisy labels and ensemble multiple predictors. The approach combines sample-focused attention with adaptive label refurbishment, achieving superior precision and average returns across stocks, cryptocurrencies, and ETFs on the Qlib platform while demonstrating robustness to label noise. Key contributions include a novel LA-Attention design with a metric learning module and a principled RA-Labeling scheme that updates labels over multiple rounds and supports Last/Vote ensembles. The results indicate meaningful practical impact for real-world quantitative trading, offering a more robust path to trading opportunities in highly noisy financial markets.

Abstract

Price movement forecasting, aimed at predicting financial asset trends based on current market information, has achieved promising advancements through machine learning (ML) methods. Most existing ML methods, however, struggle with the extremely low signal-to-noise ratio and stochastic nature of financial data, often mistaking noises for real trading signals without careful selection of potentially profitable samples. To address this issue, we propose LARA, a novel price movement forecasting framework with two main components: Locality-Aware Attention (LA-Attention) and Iterative Refinement Labeling (RA-Labeling). (1) LA-Attention, enhanced by metric learning techniques, automatically extracts the potentially profitable samples through masked attention scheme and task-specific distance metrics. (2) RA-Labeling further iteratively refines the noisy labels of potentially profitable samples, and combines the learned predictors robust to the unseen and noisy samples. In a set of experiments on three real-world financial markets: stocks, cryptocurrencies, and ETFs, LARA significantly outperforms several machine learning based methods on the Qlib quantitative investment platform. Extensive ablation studies confirm LARA's superior ability in capturing more reliable trading opportunities.

Figures (4)

• Figure 1: Illustration of the potential profitable samples. The half top figure represents the probability density function (PDF) of expected return over corresponding samples (best viewed in color).
• Figure 2: The workflow of our proposed LARA framework. LARA first extracts potentially profitable samples from the noisy market and then refines their labels. LARA consists of two sequential components: LA-Attention and RA-Labeling.
• Figure 3: Visualization of randomly sampled 1000 points with t-SNE and their corresponding return distributions. Left: the input dataset. Middle:samples with high $p_{{\bm{x}}_t}$ selected by localization module. Right:samples with high $p_{{\bm{x}}_t}$ selected by LA-Attention (localization module + metric learning module).
• Figure 4: Left: Precision with the different number of trades ($\#$Transactions) among three methods on 512480.SH. Middle: Quantitative comparisons over cumulative return between LARA and a set of baselines on 512480.SH. The right y-axis illustrates the volatility campbell1997econometrics on each trading day. Right: Plot of the trade frequency in the China's A-share market and the corresponding stock index price. The left y-axis denotes the stock index price while the right y-axis denotes the trade frequency.