AI modeling for market prediction - 1

The episode presents a research paper that introduces a novel attention-based framework for predicting stock trends by modeling both investor subjective expectations and automatically discovering dynamic latent stock topics. This approach seeks to overcome limitations in current relational stock models, which rely on predefined relationships and only consider immediate effects. Tested extensively on China’s CSI 300 market, the model consistently demonstrates an annual return exceeding 10% in trading simulations, setting a new state-of-the-art standard compared to sixteen established baseline methods across various performance metrics. The framework is composed of three jointly optimized modules: temporal stock representation, a topic module, and an expectation module. Future research aims to improve profitability stability and incorporate external factors like social media into the expectation modeling