LSTM based stock prediction using weighted and categorized financial news

Area of Science:

• Computational Finance
• Artificial Intelligence in Finance
• Natural Language Processing for Financial Markets

Background:

• Extensive research links financial news to stock market trends.
• Limited exploration of stock prediction models using weighted news categories.
• Need for models that leverage hierarchical news structures (market, sector, stock).

Purpose of the Study:

• To enhance stock prediction accuracy by integrating weighted news categories.
• To propose a novel Weighted and Categorized News Long Short-Term Memory (WCN-LSTM) stock prediction model.
• To evaluate the model's performance on the Pakistan Stock Exchange (PSX).

Main Methods:

• Development of the WCN-LSTM model incorporating learned weights for news categories.
• Integration of hybrid input, lexicon-based sentiment analysis, and deep learning for sequential learning.
• Experimentation using diverse sentiment dictionaries and time steps on the PSX dataset.

Main Results:

• The WCN-LSTM model demonstrated superior performance compared to baseline models.
• Optimal prediction accuracy was achieved using the HIV4 sentiment lexicon with time steps of 3 and 7.
• Statistical analysis confirmed the model's effectiveness and superiority over existing methods.

Conclusions:

• Incorporating weighted and categorized news significantly enhances stock prediction accuracy.
• The WCN-LSTM model offers a novel and effective approach to financial market forecasting.
• The findings highlight the importance of hierarchical news analysis in quantitative finance.