Stock Price Forecasting: Classical vs Deep Learning Models

AI in Finance Capstone Project A comprehensive comparison of ARIMA, LSTM, and GRU models for stock price prediction across three distinct market sectors.

---

📊 Project Overview

This project implements and compares three forecasting approaches for stock price prediction:

• ARIMA(3,1,5) - Classical time series model
• LSTM - Long Short-Term Memory neural network
• GRU - Gated Recurrent Unit neural network

The analysis covers three stocks with different market characteristics:

• 🍎 AAPL (Apple) - Large-cap technology, stable growth
• 🎮 NVDA (NVIDIA) - High-volatility semiconductor/AI
• 🚗 LYFT - Small-cap ride-sharing, erratic patterns

Time Period: 2020-01-01 to Present (~5 years) Data Source: Yahoo Finance Training/Test Split: 80/20

---

🎯 Key Findings

Model Performance Summary

Stock	Best Model	RMSE Improvement vs ARIMA	Directional Accuracy
AAPL	GRU	64% reduction	~52%
NVDA	LSTM	80% reduction	~51%
LYFT	GRU	80% reduction	~53%

Main Conclusions

1. Neural networks significantly outperform ARIMA for all stocks studied (64-82% RMSE reduction)
2. GRU is the practical winner - best balance of accuracy, training speed, and simplicity
3. LSTM excels for high-volatility stocks - superior long-term memory for explosive growth patterns (NVDA)
4. ARIMA remains valuable for baseline comparison and interpretable forecasts
5. Directional accuracy ~51% - even sophisticated models struggle to beat random chance (50%)

---

🗂️ Project Structure

Stock-analytics/
├── main.ipynb                              # Main analysis notebook
└── README.md                               # This file

---

📚 Milestone Breakdown

Milestone 1: Data Acquisition & Classical Models (Steps 1-7)

Completed Tasks:

• ✅ Data extraction from Yahoo Finance API
• ✅ Exploratory data analysis with interactive visualizations
• ✅ Time series decomposition (trend, seasonality, residuals)
• ✅ Stationarity testing (ADF test)
• ✅ ACF/PACF analysis
• ✅ Feature engineering (EMA, DEMA)
• ✅ Classical model implementation:
  • MA(20) - Moving Average
  • AR(20) - Autoregressive
  • ARIMA(3,1,5) - Grid search optimal model

Key Results:

• ARIMA(3,1,5) selected by grid search on NVDA (AIC = 4378.65)
• Stationarity achieved with first-order differencing (d=1)
• Applied consistently across all three stocks

---

Milestone 2: Deep Learning Models (Steps 8-9)

Completed Tasks:

• ✅ Data normalization (MinMaxScaler 0-1 range)
• ✅ Sequence generation (60-day lookback window)
• ✅ LSTM implementation:
  • 2-layer architecture (50 units each)
  • Dropout regularization (0.2)
  • Early stopping (patience=5)
• ✅ GRU implementation:
  • 2-layer architecture (50 units each)
  • 25% fewer parameters than LSTM
  • Faster training convergence
• ✅ Training/validation split (90/10 within training set)
• ✅ Convergence behavior analysis

Architecture Details:

Input: 60 days × 5 features (Open, High, Low, Close, Volume)
Layer 1: LSTM/GRU (50 units, return_sequences=True)
Dropout: 0.2
Layer 2: LSTM/GRU (50 units)
Dropout: 0.2
Dense: 25 units
Output: 1 unit (next day's closing price)

---

Milestone 3: Model Evaluation & Comparison (Steps 10-11)

Completed Tasks:

• ✅ Comprehensive metrics calculation:
  • RMSE (Root Mean Squared Error)
  • MAE (Mean Absolute Error)
  • MAPE (Mean Absolute Percentage Error)
  • DA (Directional Accuracy)
• ✅ Side-by-side model comparison (3 stocks × 3 models = 9 combinations)
• ✅ Prediction visualizations with interactive Plotly charts
• ✅ Detailed interpretation of when LSTM/GRU outperform ARIMA
• ✅ Practical recommendations for different use cases

Performance Breakdown:

AAPL (Apple):

• ARIMA(3,1,5): RMSE $24.29
• LSTM: RMSE $8.89
• GRU: RMSE $8.75 ✨ Winner

NVDA (NVIDIA):

• ARIMA(3,1,5): RMSE $42.10
• LSTM: RMSE $8.56 ✨ Winner
• GRU: RMSE $9.84

LYFT:

• ARIMA(3,1,5): RMSE $4.53
• LSTM: RMSE $1.11
• GRU: RMSE $0.92 ✨ Winner

---

🛠️ Technologies Used

Python Libraries:

• pandas - Data manipulation
• numpy - Numerical computing
• yfinance - Yahoo Finance API wrapper
• statsmodels - ARIMA, ACF/PACF, ADF test
• tensorflow/keras - LSTM/GRU implementation
• scikit-learn - Preprocessing, metrics
• plotly - Interactive visualizations

Environment:

• Python 3.8+
• TensorFlow 2.x
• Jupyter Notebook

---

🚀 Getting Started

Installation

# Clone the repository
git clone https://github.com/allanilya/Stock-analytics.git
cd Stock-analytics

# Install dependencies
pip install pandas numpy yfinance statsmodels tensorflow scikit-learn plotly

# Launch Jupyter Notebook
jupyter notebook main.ipynb

Running the Analysis

1. Open main.ipynb in Jupyter Notebook
2. Run all cells sequentially (Kernel → Restart & Run All)
3. Expected runtime: ~10-15 minutes
   • Data extraction: ~2 min
   • Classical models: ~2 min
   • Neural network training: ~5-8 min
   • Visualization: ~2 min

Requirements

pandas>=1.3.0
numpy>=1.21.0
yfinance>=0.1.63
statsmodels>=0.13.0
tensorflow>=2.8.0
scikit-learn>=1.0.0
plotly>=5.3.0
matplotlib>=3.4.0

---

📈 When to Use Each Model

Use LSTM When:

• ✅ Stock has high volatility (e.g., NVDA, tech growth stocks)
• ✅ Long-term dependencies matter (product cycles, earnings patterns)
• ✅ Computational cost is not a constraint
• ✅ You need the most accurate predictions

Use GRU When:

• ✅ Default choice for most applications
• ✅ Training speed matters (production systems)
• ✅ Stable or moderately volatile stocks (e.g., AAPL, LYFT)
• ✅ You want best accuracy-to-complexity ratio

Use ARIMA When:

• ✅ Interpretability is critical (regulatory requirements)
• ✅ Limited data available (<1000 samples)
• ✅ Quick baseline needed
• ✅ Stock follows linear patterns (index funds, utilities)

---

🔍 Limitations & Future Work

Current Limitations:

1. Univariate models - Only uses past prices, ignores volume/sentiment
2. Single architecture - Did not test 1-layer vs 3-layer networks
3. One-day-ahead only - Multi-day forecasts may favor different models
4. Train/test split - Walk-forward validation would be more rigorous
5. Directional accuracy ~51% - Barely beats random (50%)

Future Improvements:

1. Multivariate models - Add volume, RSI, MACD, sentiment scores
2. Attention mechanisms - Transformer models for sequence-to-sequence
3. Hybrid models - ARIMA for trend + LSTM for residuals
4. Architecture search - Grid search over layers, units, sequence length
5. Walk-forward validation - Rolling window retraining
6. Risk metrics - Sharpe ratio, maximum drawdown, VaR

---

📊 Sample Visualizations

The notebook includes:

• 📈 Price trends with exponential moving averages
• 🔄 Time series decomposition (trend/seasonality/residuals)
• 📊 ACF/PACF plots for model selection
• 🎯 Forecast vs actual comparison charts
• 📉 Training history (loss curves)
• 🏆 Model performance comparison tables

---

🎓 Academic Context

This project fulfills the requirements for a three-milestone capstone in AI for Finance:

• Milestone 1: Classical time series analysis with ARIMA
• Milestone 2: Deep learning implementation with LSTM/GRU
• Milestone 3: Comprehensive model evaluation and comparison

The analysis demonstrates that while neural networks significantly outperform classical models in terms of RMSE, the near-random directional accuracy (~51%) aligns with the semi-strong efficient market hypothesis - short-term stock prices are fundamentally difficult to predict.

---

📝 References

1. ARIMA Modeling: Box, G. E. P., & Jenkins, G. M. (1976). Time Series Analysis: Forecasting and Control
2. LSTM Networks: Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term Memory
3. GRU Networks: Cho, K., et al. (2014). Learning Phrase Representations using RNN Encoder-Decoder
4. Financial Time Series: Tsay, R. S. (2005). Analysis of Financial Time Series

---

👥 Authors

Allan Ilyasov Giulio Bardelli Peter Roumeliotis

---

📄 License

This project is for educational purposes as part of an academic capstone.

---

🙏 Acknowledgments

• Professor's reference notebook: 04-05 ARIMA_with_AlphaVantage_Vintage_API.ipynb
• Yahoo Finance for providing free historical stock data
• TensorFlow/Keras teams for deep learning framework
• Statsmodels for classical time series tools

---

Last Updated: 2025-11-23 Status: ✅ Complete - All three milestones fulfilled