Apple Stock Price Prediction (1980–2022) Using GRU Deep Learning

This project implements, evaluates, and deploys a deep learning–based stock price prediction system using a Stacked GRU architecture trained on Apple Inc. (AAPL) historical stock data spanning over 40 years.

The system includes a complete time series preprocessing pipeline and a production-style inference interface with multi-day autoregressive forecasting.

Live Demo: https://huggingface.co/spaces/your-username/your-space-name

---

Problem Statement

Build a regression model to predict Apple's future Close Price (USD) using historical OHLCV data, and deploy it as a real-time multi-day forecasting system.

---

Model Architecture

• MinMaxScaler Preprocessing (5 features)
• GRU Layer 1 (50 units, return_sequences=True)
• Dropout (0.2)
• GRU Layer 2 (50 units)
• Dropout (0.2)
• Dense Output Layer (Linear activation)
• Loss Function: Mean Squared Error
• Optimizer: Adam
• Look-back Window: 60 days
• Total Parameters: 23,901

The stacked GRU architecture was selected to capture long-range sequential dependencies in financial time series data.

---

Model Performance

Metric	Score
MAE	$3.34
RMSE	$6.12
MAPE	3.447%
Prediction Accuracy	96.55%

The model tracks Apple's price trajectory closely through 2014–2019. The widening gap post-2020 reflects the COVID-driven bull run — a market anomaly beyond historical training patterns, not a model deficiency.

---

Time Series Preprocessing Pipeline

The following preprocessing steps were implemented:

• OHLCV feature selection (Open, High, Low, Close, Volume)
• Min-Max scaling to [0, 1] range
• 60-day sliding window sequence creation
• 80/20 chronological train/test split
• Inverse transformation for interpretable predictions
• Full dataset retraining for production deployment

The scaler and trained model were serialized to ensure reproducible deployment.

---

Deployment & Inference System

The model is deployed using Hugging Face Spaces with:

• CSV upload interface for historical stock data
• Autoregressive multi-day forecasting (up to 30 days)
• Day-by-day price prediction table with % change
• Actual vs Forecast line chart
• Forecast summary with overall direction and range
• Stateless prediction pipeline

Model saved in .h5 format
Scaler serialized via joblib

---

Tableau Performance Dashboard

A 5-chart interactive dashboard was built to evaluate model behavior:

• Actual vs Predicted Close Price (Line Chart)
• Residuals Over Time (Bar Chart)
• Scatter Plot — Actual vs Predicted
• Model KPIs — MAE, RMSE, MAPE
• Error Distribution (Histogram)

---

Tech Stack

• Python
• TensorFlow / Keras
• Scikit-learn
• Pandas / NumPy
• Matplotlib / Seaborn
• Gradio
• Hugging Face Spaces
• Tableau Public

---

Project Highlights

• End-to-end ML pipeline from raw data to live deployment
• Reproducible model artifact serialization
• Autoregressive multi-step forecasting capability
• Interactive Tableau dashboard for model performance analysis
• Evaluation-aware design with separate test set and full dataset retraining

---

Potential Improvements

• Add technical indicators (RSI, MACD, Bollinger Bands) as features
• Implement multi-stock generalization beyond AAPL
• Add baseline comparison (ARIMA, Prophet)
• Integrate attention mechanism for interpretability
• Expose REST API using FastAPI

---

Disclaimer

This project is for educational and portfolio purposes only. Predictions are not financial advice and should not be used for investment decisions.