[MLP Project 4–2] Cross-Market Extension: KOSPI Excess Return Prediction

📌 Course Information

Category	Detail
Course	CS 53744 — Machine Learning
Instructor	Prof. Jongmin Lee
Project	MLP Project 4–2 — Cross-Market Extension
Objective	Extend the S&P 500 forecasting pipeline to a new market (KOSPI)

---

📌 Team Information

Role	Name (KOR)	Student ID	GitHub ID
Member	박원규	20203955	@keiro23
Member	이유정	20214677	@yousrchive
Member	정승환	20201799	@whan0767

---

🎯 Project Goal

This repository contains a fully reproducible implementation of the Cross-Market Extension bonus project.
The goal is to adapt the Hull Tactical S&P 500 prediction pipeline to a different financial environment —
specifically, the KOSPI index — and evaluate:

1. Whether the short-horizon predictability found in the U.S. market transfers to Korea
2. How market microstructure differences affect ML-based forecasting
3. How to build a volatility-constrained allocation strategy using the predicted excess returns

Target variable:

market_forward_excess_returns — next-day excess return relative to a 5-year rolling mean

Allocation rule:

w_t = clip(1 + k*z_t, 0, 2) with volatility ≤ 1.2× benchmark

---

Repository Structure

proj4-2/
├── data/
│   ├── 뉴스심리지수.xlsx                 # BOK Snapshot에서 수동 다운로드한 NSI
│   ├── kospi_df.csv                      # 두 데이터를 머지하고 feature engineering된 KOSPI 데이터
│   └── kospi_yahoo_AB_full_feature.csv # 수집한 원본 데이터
│
├── src/
│   ├── 01_data_preprocessing.ipynb       # 데이터 로딩, 정제, NSI merge, feature 생성
│   ├── 02_EDA.ipynb                      # 탐색적 분석, 시장 구조 분석
│   ├── 03_Modeling.ipynb                 # PCA + ElasticNet 모델링 + Backtesting
│   └── requirements.txt
│
└── README.md         

---

3. Dataset Card

Data Components

Category	Columns	Source	Notes
Price Data	date, open, high, low, close, volume	Yahoo Finance (^KS11)	via yfinance
Returns / Momentum	ret_1d, ret_5d, ret_22d, momentum_10/20/60	Derived	Sliding-window calculations
Trend Indicators	ma_* , MACD, signal, histogram	Derived	Captures regime behavior
Volatility / Risk	vol_22d, ATR, Bollinger Bands, RSI	Derived	Reflects uncertainty
Drawdowns	drawdown_60, max_drawdown_60	Derived	Downside risk behavior
External Indicators	VIX, USD/KRW, Gold price	Yahoo Finance	Global risk signals
News Sentiment Index	NSI	BOK Snapshot Excel download	Added manually

Collection Notes

• All financial price series collected automatically with yfinance.
• NSI (뉴스심리지수.xlsx) was manually downloaded from: https://snapshot.bok.or.kr/dashboard/C8
• The dataset contains all raw and engineered features required for modeling.

---

4. Modeling Pipeline

✔ PCA (Dim = 5)

• Explained ~85% of variance.
• PC1 = market trend
• PC2 = volatility regime
• PC5 = sentiment + FX shock loading (Korea-specific)

✔ ElasticNet on PCA

• Best-performing linear model
• Nonlinear models (LightGBM/XGB) overfit KOSPI’s high-noise structure

✔ Time-Series Cross-Validation

• 5-fold TimeSeriesSplit
• PCA fit only on train split → leakage-free

✔ Volatility-Constrained Allocation

z_t = standardize(pred_t)
w_t = clip(1 + k·z_t, 0, 2)
scale if volatility exceeds 1.2× benchmark

---

5. Results Summary

Test Out-of-Sample (OOS) Results

Metric	Benchmark	PCA+ElasticNet
Annualized Return	–6.7%	+2.8%
Annualized Vol	16.4%	19.7%
Vol Ratio	1.00	1.20 (capped)
Sharpe	–0.41	0.14
Max Drawdown	–39.2%	–33.0%

Interpretation

• Predictability in KOSPI exists but is very weak.
• However, PCA+ElasticNet extracts small signals that translate into a Sharpe improvement under a volatility cap.
• NSI & USDKRW contribute to Korea-specific factors → not present in S&P 500.
• Nonlinear models degrade performance due to overfitting.

---

6. How to Run Everything

Install Dependencies

pip install -r src/requirements.txt

Run the workflow

1. 01_data_preprocessing.ipynb → Load Yahoo Finance price data → Merge NSI → Create full feature dataset

2. 02_EDA.ipynb → Plot rolling stats, vol regimes, sentiment correlation

3. 03_Modeling.ipynb → PCA, ElasticNet, time-series CV → Backtest (cumulative return / drawdown / rolling vol) → Export results figures for report appendix

All notebooks produce reproducible outputs using only files inside data/.

---

7. Compliance with Course Requirements

✔ Dataset card (included above)

✔ Reproducible source code (in src/)

✔ Figures & results for Appendix (generated in notebook 03)

✔ Volatility-constrained strategy

✔ EMH interpretation & qualitative discussion

✔ No external API requiring credentials

✔ No leakage in modeling (verified via PCA per-fold training)

---

8. License & Disclaimer

None of the models, signals, or strategies should be used for real-world investment.