Using-Multimodal-Model-AI-to-Develop-Stock-Trading-Strategy

Main Points

Preface

• Objective: Improve the accuracy of portfolio return predictions to help investors make better decisions, increase returns, and reduce risks.
• Technologies Used: Implemented advanced machine learning techniques including Graph Neural Networks (GNN) to handle the complexity and variability of financial markets.

Problem

• Three Main Analysis Directions:
  • Fundamental Analysis: Based on financial reports, updated monthly, with limited data for high-frequency analysis.
  • Sentiment/Flow Analysis: Includes market sentiment and capital flows.
  • Technical Analysis: Relies on market data like stock prices and volumes to predict future trends.
• Challenge: Traditional methods focus on single-company data, neglecting the impact of supply chain or competitor activities on a company's stock price.

Approach

• Combined Analysis: Integrated sentiment (news) and technical (candlestick charts) analysis using GNN to predict portfolio returns, considering upstream, downstream, and competitive relationships among companies.
• Data Integration: Utilized multiple data sources to improve prediction accuracy, forming portfolios and backtesting overall prediction accuracy.

Project Goal

• Target Stocks: Selected top 60 companies by market capitalization in Taiwan from 2014.
• Goal: Predict stock movements to build portfolios, based on the assumption of market inefficiency, integrating various forms of information to capture risk factor-driven returns.

News Text Processing

• Three Parts:
  • Data Collection: Web scraping for relevant stock news.
  • Text Analysis: Used transformer-based language models to extract financial features from news.
  • Integration with Other Data: Combined text features with financial indicators for comprehensive analysis.
• Data Collection: Gathered news content for 58 stocks from 2014 to 2024, with predictions made every 12 days.
• Modeling: Used a Chinese BERT model for preprocessing news data, applying transformer architecture for accurate textual representation.

Stock Chart Vector Processing

• Data Collection: Used Yahoo Finance API to obtain OHLC and volume data from 2013.7.1 to 2024.7.20.
• Technical Analysis: Added 20-day moving average (MA) to avoid excessive noise in charts.
• Labeling: Labeled data as 1 if the price increased after 10 days, otherwise 0.
• Image Generation: Created grayscale candlestick charts using numpy arrays, covering 60 trading days with normalized data.
• CNN Training: Trained a 3-layer CNN with 5x3 convolution size, outputting vectors for GNN training.
• Additional Models: Experimented with VGGNet, ResNet, and U-net.

Macroeconomic Indicator Vector

• Objective: Provide the model with a broader view by integrating macroeconomic data, which is consistent across all stocks.
• Data Collection: Used Yahoo Finance API and FRED API to obtain relevant indicators.
• Indicators: Included exchange rates, volatility, commodity prices, CPI inflation, unemployment rates, interest rates, and money supply.
• Normalization: Performed 60-day normalization to ensure consistency.
• GNN Training: Passed the data to GNN for training.

GNN (Graph Neural Network)

• Stock Price Influences: Considered not just company-specific data but also impacts from suppliers, customers, and competitors.
• GNN Models:
  • Total Network Graph: Includes all competitive relationships.
  • Heterogeneous Graph: Focuses only on competitive relationships.
  • Homogeneous Graph: Includes only cooperative relationships, such as those between suppliers and customers.
• Node Creation: Developed a new node related to all companies, using the Taiwan stock market index to represent overall market conditions.