• Input Features
• Key Features
• Target Selection Strategy
• Model Hyperparameters
• Labeling & Ranking
• Return Evaluation
• Case Study
• Environment
• Experiment Notebook
• License & Acknowledgements

A chronological overview of model improvements, from LSTM to Transformer variants.

📘 See TRANSFORMER.md for a detailed architecture explanation.

Patches are extracted via Conv1D to compress temporal data, and a Transformer encoder captures inter-patch dependencies for ranking prediction.

📘 Full feature list: FEATURES_TECHNICAL_INDICATORS.md

• Feature engineering:
  Ratio-based transformations, log-scaling for long-tail features, stable across different stocks.

• industry_id, is_kospi

• close_rate, open_to_close, high_to_low, rsi, ato, macd related

• vwma5_gap, vwma20_gap, vwma_bb_width etc.

• KOSPI, KOSDAQ, S&P500 Futures, Nasdaq 100 Futures, VIX, etc.

• Volume volatility ratios, net buy rates

• Upper/lower tail ratios, body ratios

The following heatmap shows pairwise correlations between selected input features.

Below are the individual distributions of input features used in model training.

• Patch-wise Transformer encoder
• Industry embedding
• Soft label generation from 5-day future returns
• LambdaRankLoss for ranking optimization
• Real-time applicability (15:40–16:00 trading window)

• Top 200 stocks by daily trading volume
• Market cap ≥ 500B KRW
• Excludes limit-up and newly listed stocks

• Input Dim: 41
• Industry Embedding Dim: 4
• Model Dim: 64
• Sliding Window Size: 30
• Patch Length: 6, Stride: 3
• Transformer Encoder Heads: 4, Layers: 2
• Dropout: 2
• Learning Rate: 5e-4
• Weight Decay: 5e-5
• Early Stopping Patience: 10 epochs

• 5-day return quintiles (20 bins) used for soft labels
• TOP3 selection by predicted score per day

For an in-depth explanation of the ranking metric and training objective used in this model, see:

• 📘 NDCG Explained:
  Understand how Normalized Discounted Cumulative Gain (NDCG) measures ranking quality in stock selection.

• ⚙️ LambdaRank Loss Guide:
  Dive into the pairwise ranking loss function that optimizes NDCG by comparing stock relevance in every batch.

While model prediction provides initial candidates, additional filtering and dynamic sell strategies are applied to make the system robust for real-world trading.

• pred_rank <= topn (e.g., TOP3)
• atr_ratio > 0.03 (sufficient volatility)
• close_rate > -10% (avoiding sharp decliners)

• Max holding period: Forced exit after fixed days (e.g., 5 days).
• Trailing Entry Extension:
  If a new buy signal occurs during holding, reset holding period.

• Python 3.12.8
• PyTorch 2.6.0 + CUDA 12.6
• pykrx 1.0.48
• Full list in requirements.txt

The full end-to-end workflow is implemented in the following notebook:
Stock_PatchTST_Ranking.ipynb

This includes:

• Raw data retrieval
• Feature engineering and preprocessing
• Model training and validation
• Return evaluation and analysis
• Case studies on selected stocks

This project is licensed under the MIT License.
See the LICENSE file for details.

This work is inspired by PatchTST.
It was developed and tested on KRX daily stock data from 2020 to 2025. KRX data was primarily retrieved using the pykrx library.
Industry classification codes were retrieved via the Open API from Korea Investment & Securities.