Stock Price Prediction: PatchTST vs Baseline Models

Comparative study of deep learning architectures for S&P 500 stock price forecasting.

Models

Baseline (stock_prediction_baseline.ipynb)

• MLP (2-layer, 256 hidden units)
• CNN (1D Conv with 64→128→256 channels)
• LSTM (2-layer, 256 hidden, bidirectional)

PatchTST (stock_prediction_patchtst.ipynb)

• Transformer-based architecture with patch tokenization
• RevIN (Reversible Instance Normalization) for distribution shift
• Config: 64 lookback, patch_len=8, stride=4, d_model=384, 8 heads, 5 layers

Features

Technical indicators derived from OHLCV data:

• Momentum: RSI, MACD, ROC
• Volatility: ATR, Bollinger Bands
• Volume: OBV, VWAP
• Trend: ADX, EMA

Dataset

S&P 500 historical data (5 years) from Kaggle.

Usage

pip install torch pandas pandas_ta matplotlib seaborn

Run notebooks in order:

1. stock_prediction_baseline.ipynb - Train baseline models
2. stock_prediction_patchtst.ipynb - Train PatchTST with weighted ensemble

Metrics

• RMSE (Root Mean Square Error)
• MAPE (Mean Absolute Percentage Error)
• Directional Accuracy
• R² Score

Authors

• Mohammad Farid Hendianto (2200018401)
• Fidyah Rahman (2200018185)

Kapita Selekta - Kelompok 3

Report

📄 Read Full Academic Report (PDF)
Comprehensive analysis of PatchTST vs Baseline models, including methodology, architectural diagrams, and experimental results.

License

Copyright © 2026 Mohammad Farid Hendianto (IRedDragonICY).

Licensed under the Apache License, Version 2.0. See LICENSE for more details.