AIAI Seminar-27 May-Talk by Mengyu Wang and Hao Zhou

Speaker: Mengyu Wang

Title: FINRank: Financial Influence News Ranking for Better Market Prediction

Abstract: Financial news is recognized as a significant indicator of gaining insight into market dynamics. However, assigning weights to news items based on their potential market impact remains a challenge. The intricate and often elusive connections between financial news and market movements complicate the ranking annotation of financial news datasets. This paper presents a novel approach for training a financial news ranking model. We introduce FINRank (Financial Influence News Ranking) model, which is designed to generate news ranking scores through the analyze of multi-modal information from both the stock market and news items. By embedding FINRank into a stock prediction framework, we convert the training objective from news ranking to market prediction. This allows us to effectively train a transferable news ranking model without relying on ranking labels. Experiments conducted on two comprehensive datasets, including the S&P500 index spanning 15 years and over 35 million news articles, illustrate FINRank's effectiveness. Our news ranking scores exhibit the ability to improve risk-adjusted returns, outperforming various recent baselines with an improvement of 0.50 in the average Sharpe ratio. Furthermore, our results reveal several aspects of the interrelation between the market and news, including the lagged relationships between market pricing and news, the news long-term impact, and the limitations of financial news sentiment analysis. These insights contribute to a more profound understanding of the market information diffusion process.

Speaker: Hao Zhou

Title: FeatureNet: A Multiple Dimensional Features and Temporal Dependencies Model for Stock Price Movement Prediction

Abstract: Stock price movement prediction is inherently a challenging task, as future prices are a function of the complex interrelationships between features, where causality and the strength of relationships between such features and their effects on stock prices can fluctuate over time. Common predictive models in this domain often experience two key limitations: they treat each time step in isolation, neglecting cross-time dependencies, or they overlook the multivariate and nonlinear nature of cross-feature dependencies. This study introduces FeatureNet, a Transformer-based model for predicting the S&P 500 price, which explicitly addresses cross-time and cross-feature dependencies arising from common features on stock prices, namely, technical, fundamental, and sentiment features. Comparative analysis against established models demonstrates our FeatureNet’s significant improvements over four distinctive metrics of prediction power and calibration, thus indicating the importance of accounting for the cross-time and cross-feature relationships on stock price movement prediction.