ECAWE: Stock Price Predictions based on Extreme Cross Attention and Weighted Extreme Loss Function

Stock prices exhibit complex dynamics influenced by political events, economic cycles, and investor sentiment, posing substantial challenges to accurate forecasting. A particularly challenging aspect is the non-stationarity and heavy-tailed distribution of price series, where empirical analysis shows extreme values (exceeding the 80th percentile) account for over 40% of the total mean squared error (MSE) in conventional machine learning models. To address this, we propose a novel framework incorporating an extreme cross-attention (ECA) mechanism that explicitly captures the interplay between normal patterns and extreme events. Our approach: (1) processes multi-sector stock data to extract multidimensional features (e.g., technical indicators); (2) identifies extreme values using quantile thresholds of stock price sequences; (3) computes bidirectional attention weights between normal-state and extreme-value embeddings via ECA; (4) optimizes predictions through a hybrid loss function that combines Huber loss (for normal fluctuations) and a Fréchet-based extreme loss (for extreme points). Extensive experiments across seven industry stock datasets demonstrate that our model reduces MSE by 17-36% against both classical (e.g., LSTM) and state-of-the-art deep learning benchmarks (e.g., iTransformer). Ablation studies confirm the contribution of each component, and a portfolio case study demonstrates its practical utility.