Stock Price Prediction Application

Inspiration

This project was inspired by the challenge of predicting stock prices in a volatile market, particularly for high-growth or speculative stocks like CELH, CVNA, UPST, ALT, and FUBO. I wanted to explore how combining technical indicators, fundamental data, macroeconomic factors, and news sentiment could improve forecasting accuracy. The idea stemmed from a desire to create a practical tool for investors or traders, blending my interest in machine learning with real-world financial applications.

What I Learned

• Hybrid Modeling: I learned to integrate LSTM (for time-series data) with dense layers (for static fundamental features) using PyTorch, creating a HybridModel that leverages multiple data types.
• Feature Engineering: Calculating technical indicators (e.g., RSI, MACD, Bollinger Bands) and incorporating sentiment analysis from news taught me the importance of diverse feature sets.
• API Integration: Working with yfinance and fredapi showed me how to fetch and align real-time financial and economic data.
• Streamlit: Building an interactive dashboard was my first hands-on experience with Streamlit, making complex predictions accessible to users.
• Challenges of Stock Prediction: The mixed accuracy results (e.g., 90% for UPST vs. 14% for CVNA) highlighted the unpredictable nature of stock markets and the limits of historical data alone.

How I Built It

1. Data Collection:

   • Used yfinance to download stock price data and fundamentals (e.g., PE ratio, debt-to-equity) for a specific period (Jan 21, 2024 - July 21, 2024).
   • Fetched macroeconomic data (interest rates, inflation) via fredapi.
   • Extracted news sentiment using yfinance’s news feature, processed with TextBlob and VADER.

2. Feature Engineering:

   • Calculated technical indicators like moving averages, volatility, RSI, and MACD using Pandas.
   • Added macroeconomic features (e.g., interest rates, VIX) and sentiment scores to enrich the dataset.

3. Modeling:

   • Built a HybridModel in PyTorch combining LSTM (for technical time-series) and dense layers (for fundamentals).
   • Trained the model on sequences of technical data (30-day windows) and static fundamental features, using MSE loss and Adam optimizer.
   • Saved trained models to disk for reuse.

4. Prediction & Visualization:

   • Generated predictions with the trained HybridModel and saved them as CSV files.
   • Created a Streamlit dashboard to let users select stocks, trigger predictions, and view results with Plotly charts.

5. Structure:

   • Modularized the code into data_loader.py (data fetching), features.py (feature engineering), models.py (model definitions), train.py (training), and predict.py (inference), with a main script to tie it all together.

Challenges Faced

• Data Alignment: Ensuring price data, news, and macroeconomic indicators aligned correctly over the date range (Jan 21 - July 21, 2024) was tricky. I used Pandas’ reindex and forward-fill to handle gaps.
• Model Accuracy: The huge variance in prediction accuracy (e.g., CVNA at 14.7% vs. UPST at 90.1%) suggested overfitting or missing key features for volatile stocks. CVNA’s wild deviation was a puzzle—possibly due to a market event my model didn’t capture.
• Column Naming: The Streamlit app crashed due to a KeyError: 'predicted' because my predict_stock function named the output column prediction. I had to adapt the app to handle flexible column names.
• Feature Availability: Some fundamental data (e.g., PE ratio) was missing for certain stocks, requiring imputation with defaults, which might have skewed predictions.
• Computational Resources: Training the hybrid LSTM model for multiple stocks with 100 epochs was resource-intensive, pushing me to optimize batch sizes and validation splits.

Next Steps

• Improve CVNA/CELH Predictions: Investigate market events or add more features like social media sentiment (e.g., X posts) to capture volatility.
• Ensemble Refinement: Weight models by performance or integrate Prophet/RandomForest more effectively with the LSTM.
• Real-Time Data: Extend the app to fetch live data for ongoing predictions.
• User Experience: Enhance the Streamlit UI with date range selectors and more detailed error messages.

This project was a deep dive into financial machine learning, blending theory with practical application. While the results aren’t perfect, they’ve given me a solid foundation to iterate and improve!

Built With

• bollinger-bands
• csv-predictions
• dense-(fundamentals)
• ensemble-learning.-feature-engineering:-ma
• file
• fred
• hybrid-model
• languages-&-frameworks:-python
• macd
• macroeconomic-factors
• manual
• nan-handling-(ffill())
• newsapi-(placeholder).-architecture:-lstm-(technical)
• nltk
• no-model-version-control
• numpy
• pandas
• placeholder-sentiment
• prophet
• pytorch
• random-forest
• rsi
• scikit-learn
• sentiment-analysis.-infrastructure:-.pth-model-files
• textblob.-data-sources-&-apis:-yahoo-finance
• venv-for-dependencies.-technical-debt:-limited-news-api