TickerCast is a deep learning project that predicts next-day stock movement using a combination of financial news headlines and technical market indicators. The goal is to explore whether publicly available textual and numerical signals can inform short-term market direction and support simple trading strategies.
- Objective: Predict whether GOOGL stock will go up or down the next day using daily financial news and technical indicators.
- Approach: Combine natural language processing (news headlines) with quantitative time-series data (price-based indicators) using a multimodal neural network.
- Extras: The project also integrates a small language model (Flan-T5) to generate actionable investment insights on high-volatility days.
- News Dataset: Google financial news headlines from 2000–2023 (Kaggle)
- Price Data: GOOGL open/close/volume data fetched from Stooq via
pandas_datareader
- Cleaned and tokenized news headlines
- Converted timestamps and merged with daily stock data
- Engineered features: 5-day and 20-day moving averages (MA5, MA20), volume % change
- Baseline Model: TF-IDF + Dense network on headlines
- Advanced Model: Multimodal neural network with:
- Text branch: TextVectorization + Embedding + Conv1D
- Numeric branch: MA5, MA20, and volume change
- Fusion layer and final prediction via sigmoid output
- Classification report, confusion matrix, and ROC curve
- Strategy backtesting: Buy if
P(up) > threshold, else hold - Comparison against buy-and-hold strategy using equity curves
- Integrated
Flan-T5(via Hugging Face) to simulate a financial analyst - Generated bullet-point investment insights on days with large predicted drops
- The baseline model achieved ~66% accuracy with a ROC AUC ~0.7
- The multimodal model provided better separation of market signals
- Backtesting showed the model-based strategy avoided major losses during bearish periods
- The integrated LLM added interpretability by providing human-like explanations
├── data/ # Raw and processed datasets ├── models/ # Saved model checkpoints ├── notebooks/ # Jupyter notebooks for EDA, modeling, and backtesting ├── llm/ # Flan-T5 integration and prompt code ├── utils/ # Helper functions and preprocessing scripts └── README.md # Project overview
- Python, TensorFlow/Keras
- pandas, NumPy, scikit-learn
- Matplotlib for visualization
- Hugging Face Transformers (Flan-T5)
- Google Colab for development environment
- Multimodal modeling: fusion of text and time-series indicators
- Realistic trading simulation and return comparison
- AI-generated investment commentary
- Clean, explainable notebook workflow