Stock Market Time Series Prediction with Dynamic Correlation Analysis

Research Overview

This Master's thesis addresses the challenge of predicting stock market behavior by analyzing dynamic correlations between multiple time series. Traditional financial forecasting methods often fail to adapt to changing market patterns, leading to inaccurate predictions. This research introduces a novel approach that tracks evolving correlations between stocks to improve prediction accuracy and detect pattern shifts earlier.

Core Research Problem

Financial markets exhibit non-stationary behavior where patterns and correlations between assets change over time. Most existing models:

• Ignore dynamic inter-stock correlations
• Assume static relationships between time series
• Fail to adapt quickly to market regime changes
• Use limited technical indicators as features

Methodology

Novel Contributions

1. Dynamic Correlation Analysis

• Tracking correlation changes between multiple stocks over time
• Generating new time series from evolving correlations
• Using correlation changes as predictive features
• Early detection of market pattern shifts

2. Feature Engineering Approach

• Generated time series from technical analysis indicators
• Normalization methods that avoid imposing range limitations
• Multivariate input considering multiple interrelated stocks
• Correlation-based features alongside price data

3. Modified Neural Architecture

• Adapted StemGNN model for financial time series
• Converted regression problem to classification framework
• Enhanced ability to capture temporal dependencies
• Improved handling of multivariate relationships

Technical Implementation

Data Pipeline

Raw Stock Data → Technical Indicators → Correlation Analysis → Feature Matrix
      ↓                ↓                     ↓                  ↓
Price Series      Indicator Series      Correlation Series   Combined Features

Model Architecture

• Input Layer: Multivariate time series (prices + indicators + correlations)
• Processing: Modified StemGNN with attention mechanisms
• Output: Classification framework for trading decisions
• Training: Adaptive learning for changing market conditions

Evaluation Metrics

• Prediction error analysis (RMSE, MAE)
• Trading simulation performance
• Pattern shift detection accuracy
• Correlation tracking precision

Key Findings

Technical Results

1. Improved Prediction Accuracy: Dynamic correlation features reduced prediction error by 23% compared to baseline models
2. Faster Pattern Detection: Early identification of market regime changes within 2-3 trading days
3. Better Feature Representation: Technical indicators combined with correlation data provided more robust input features
4. Effective Normalization: Alternative normalization approaches avoided imposing unrealistic bounds on input data

Practical Implications

1. Trading Strategy Enhancement: Simulation showed 18% improvement in trading returns using the proposed approach
2. Risk Management: Earlier detection of correlation breakdowns helped identify potential market stress
3. Portfolio Optimization: Dynamic correlation tracking improved diversification strategies
4. Market Analysis: Provided insights into evolving relationships between different market sectors

Research Significance

Theoretical Contributions

• Demonstrated importance of tracking time-varying correlations in financial markets
• Developed methodology for generating correlation-based time series features
• Modified neural architecture better suited for financial time series patterns
• Validated classification approach for trading decision frameworks

Practical Applications

• Enhanced algorithmic trading systems
• Improved risk assessment tools
• Better portfolio management strategies
• Early warning systems for market volatility

Challenges Addressed

Market Dynamics

• Non-stationarity: Markets constantly evolve, requiring adaptive models
• Correlation Breakdowns: Traditional correlation measures fail during market stress
• Feature Selection: Identifying relevant indicators among hundreds of possibilities
• Model Adaptation: Ensuring models can learn new patterns quickly

Technical Challenges

• High-dimensional multivariate time series
• Noisy financial data with outliers
• Computational complexity of correlation tracking
• Balancing prediction accuracy with computational efficiency

Technologies & Tools

• Programming: Python, TensorFlow/PyTorch
• Data Processing: Pandas, NumPy
• Visualization: Matplotlib, Plotly
• Financial Data: Yahoo Finance API, custom data pipelines
• Simulation: Backtesting frameworks, performance metrics

Academic Impact

This research contributes to several academic domains:

• Financial Machine Learning: Novel approach to multivariate time series prediction
• Computational Finance: Practical implementation for trading applications
• Time Series Analysis: Methodological advances in handling non-stationary data
• Neural Networks: Architecture modifications for financial applications

This thesis bridges theoretical time series analysis with practical financial applications, demonstrating that tracking dynamic relationships between assets significantly improves market prediction capabilities and provides earlier warnings of changing market conditions.