📚 Table of Contents

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

Overview of Deep Learning Model Evolution

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

Transformer-Based Ranking Model for Stock Selection

📘 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.

---

🧠 Input Features

📘 Full feature list: FEATURES_TECHNICAL_INDICATORS.md

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

🏭 Stock Metadata

• industry_id, is_kospi

💹 Price Flow Indicators

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

📊 VWMA & Bollinger Bands

• vwma5_gap, vwma20_gap, vwma_bb_width etc.

🌍 Market Indices

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

🔁 Volume & Flow

• Volume volatility ratios, net buy rates

📈 Candlestick Patterns

• Upper/lower tail ratios, body ratios

Feature Heatmap

The following heatmap shows pairwise correlations between selected input features.

Distribution per Feature

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

---

🔧 Key Features

• 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)

🎯 Target Selection Strategy

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

🔧 Model Hyperparameters

• 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

🏷️ Labeling & Ranking

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

🧮 20-Quantile Label Bins

📉 Distribution of Predicted Scores

🔍 Raw Label Distribution

📚 Learn More: Ranking Metrics & Loss

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.

---

🔍 Post-Filtering Rules

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

🛒 Buy Signal Post-Filtering

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

---

💵 Sell Signal Detection

• 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.

---

📈 Return Evaluation

Item	2024 TOP3	2025 TOP3
Number of Samples	195,866	42,966
Number of Buy / Sell Trades	205 / 205	35 / 35
Win Rate (Count, Ratio)	112 trades (54.63%)	26 trades (74.29%)
Loss Rate (Count, Ratio)	93 trades (45.37%)	9 trades (25.71%)
Average Return (Win)	6.23%	7.54%
Average Return (Loss)	-4.57%	-3.88%
Avg. Holding Period (Win)	9.6 calendar days (6.4 trading days)	9.2 calendar days (5.8 trading days)
Avg. Holding Period (Loss)	9.5 calendar days (6.5 trading days)	10.7 calendar days (6.2 trading days)
Return Deciles	[-36.0, -6.5, -3.6, -2.0, -0.7, 0.8, 2.1, 4.1, 5.9, 9.9, 34.6]	[-8.2, -3.8, -2.3, 0.5, 1.1, 2.5, 4.2, 7.2, 8.3, 16.3, 30.7]
Trade Capital	10,000,000	10,000,000
Expected Net Return	1.335%	4.601%
Cumulative Net Profit	27,365,419	16,105,142

📅 Monthly Return

📊 Daily Return

🏦 Return by Stock

🤔 Return Distribution by Purchase Decision

🔍 Case Study: Specific Stocks

---

💻 Environment

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

🧪 Experiment Notebook

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

---

📄 License & Acknowledgements

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.