zi-gemini: Cross-Sectional Stock Market Dataset

Quantitative stock market dataset covering US, KOSPI (KO), and KOSDAQ (KQ) exchanges. Designed for cross-sectional stock return prediction and portfolio construction research.

Dataset Summary

Exchange	Stocks	EOD Rows	Date Range	Features
US	24,481 (universe: 1,685)	26.4M	2018-01 ~ 2026-05	92
KO (KOSPI)	2,401 (universe: 428)	2.6M	2018-01 ~ 2026-05	92
KQ (KOSDAQ)	1,946 (universe: 587)	3.5M	2018-01 ~ 2026-05	92

Total: 32.5M EOD rows, 48,828 tickers, 8+ years of daily data.

Data Sources

• EOD prices: Daily OHLCV with adjusted close (split/dividend adjusted)
• Fundamentals: Market cap, valuation ratios, profitability, ownership
• Earnings: Historical EPS actual/estimate/surprise, quarterly financials
• Macro: FRED (VIX, 10Y yield, DXY, WTI), BOK (KRW, base rate, CPI, bonds), Fama-French 5-factor + momentum
• Alternative: News sentiment, analyst ratings, insider transactions, economic events calendar
• Sector: KRX official classification (KO/KQ), GICS sectors (US)

Directory Structure

data/
├── KO/                          # KOSPI
│   ├── eod.parquet              # Daily OHLCV (2.6M rows, 53MB)
│   ├── fundamentals.parquet     # Company fundamentals (1,451 stocks)
│   ├── earnings_history.parquet # EPS actual/estimate/surprise (25K rows)
│   ├── quarterly_financials.parquet # Revenue/profit (20K rows)
│   ├── dividends.parquet        # Dividend history (3K rows)
│   ├── krx_sectors.parquet      # KRX official sectors
│   ├── ipos.parquet             # IPO calendar
│   ├── tickers.parquet          # Stock metadata
│   └── universe.parquet         # Investable universe (428 stocks)
│
├── KQ/                          # KOSDAQ (same structure as KO)
│   ├── eod.parquet              # 3.5M rows, 69MB
│   ├── ...
│   ├── news_sentiment.parquet   # News sentiment scores
│   └── sentiment_daily.parquet  # Daily aggregated sentiment
│
├── US/                          # US stocks
│   ├── eod.parquet              # 26.4M rows, 461MB
│   ├── fundamentals.parquet     # 18,983 stocks with 25+ fields
│   ├── earnings_history.parquet # 131K earnings reports
│   ├── quarterly_financials.parquet # 158K quarterly reports (1985~)
│   ├── analyst_ratings.parquet  # Analyst consensus (1.5K stocks)
│   ├── insider_transactions.parquet # Insider buy/sell (15K txns)
│   ├── news_sentiment.parquet   # News sentiment (41K articles)
│   ├── sentiment_daily.parquet  # Daily sentiment (244K rows)
│   ├── dividends.parquet        # Dividend history (28K rows)
│   ├── ipos.parquet             # IPO calendar
│   ├── tickers.parquet          # 24,481 tickers
│   └── universe.parquet         # Investable universe (1,685 stocks)
│
├── macro/                       # Macroeconomic data
│   ├── macro_raw.parquet        # FRED + BOK + Fama-French (36K rows, 1960~)
│   └── economic_events.parquet  # GDP/PMI/CPI events (US/KR/JP/CN)
│
└── features/                    # Pre-computed ML features
    ├── US_features.parquet      # 92 features x 1.2M rows (608MB)
    ├── KO_features.parquet      # 92 features x rows (67MB)
    ├── KQ_features.parquet      # 92 features x rows (124MB)
    ├── US_predictions.parquet   # Walk-forward OOS predictions
    ├── KO_predictions.parquet
    └── KQ_predictions.parquet

Table Schemas

eod (Daily OHLCV)

Column	Type	Description
code	string	Ticker symbol
exchange	string	Exchange code (KO/KQ/US)
date	date	Trading date
open	float	Open price
high	float	High price
low	float	Low price
close	float	Close price
adjusted_close	float	Split/dividend adjusted close
volume	int	Trading volume

fundamentals

Column	Type	Coverage	Description
market_cap	float	91%	Market capitalization
pe_ratio	float	47%	Price/Earnings ratio
pb_ratio	float	100%	Price/Book ratio
dividend_yield	float	36%	Dividend yield
roe	float	100%	Return on equity
forward_pe	float	100%	Forward P/E
price_sales_ttm	float	100%	Price/Sales TTM
ev_ebitda	float	100%	EV/EBITDA
profit_margin	float	100%	Net profit margin
operating_margin	float	100%	Operating margin
roa	float	100%	Return on assets
peg_ratio	float	44%	PEG ratio
beta	float	88%	Beta
week52_high	float	100%	52-week high
week52_low	float	100%	52-week low
shares_float	float	100%	Shares float
pct_insiders	float	100%	Insider ownership %
pct_institutions	float	100%	Institutional ownership %
sector	string	97%	GICS sector
industry	string	97%	GICS industry

features (92 pre-computed features)

Organized into categories:

• Momentum (8): ret_5d, ret_20d, ret_60d, ret_120d, kama_ratio, kama_slope, ret_5d_sector_neutral, ret_20d_sector_neutral
• Volume (3): volume_ratio_20d, obv_slope_20d, mfi_14
• Volatility (4): realized_vol_20d, realized_vol_ratio, atr_ratio, realized_skew_20d
• Technical (2): rsi_14, bb_position
• Fundamental (12): ep_ratio, bp_ratio, sales_yield, div_yield, roe, debt_to_equity, sales_growth_yoy, forward_ep, ev_ebitda_inv, profit_margin, peg_inv, roa
• Earnings & Analyst (7): earnings_surprise, revenue_surprise, analyst_target_upside, analyst_rating_norm, insider_net_flow, news_sentiment_avg, revenue_growth_qoq
• Sector Momentum (5): sector_ret_5d, sector_ret_20d, sector_ret_60d, sector_strength, stock_vs_sector
• AlphaGen RL (4): ag_delayed_mom_rank, ag_log_volume_60d, ag_open_rank_delta, ag_return_ma_delta
• Macro (18): vix, us10y, dxy, wti, usdkrw, bok_rate, cpi, term_spread, credit_spread, Fama-French 5+MOM
• Interaction (4): mom_x_value, rsi_x_vol, quality_x_mom, flow_x_value
• Market breadth (1): kospi_breadth
• Target: target (z-scored 5-day forward return), fwd_ret (raw 5-day forward return)

Universe Filtering

The universe table contains investable stocks filtered by:

• Common Stock only (no ETF/ETN/preferred)
• No SPACs
• Market cap > 20th percentile
• 20-day average dollar volume > $10M (US) / 1B KRW (KO/KQ)
• Listed > 120 trading days

Model Performance (Walk-Forward OOS)

Using 8-model ensemble (LightGBM, XGBoost, Ridge, MLP, Transformer, TabNet, StockMixer, DoubleAdapt):

Metric	US
Mean IC	0.019
Sharpe (net)	1.36
Annual Return (net)	22.0%
Max Drawdown	8.4%
Calmar Ratio	2.64

Usage

import polars as pl

# Load US EOD data
eod = pl.read_parquet("data/US/eod.parquet")

# Load pre-computed features
features = pl.read_parquet("data/features/US_features.parquet")

# Load fundamentals
fund = pl.read_parquet("data/US/fundamentals.parquet")

# Load macro data
macro = pl.read_parquet("data/macro/macro_raw.parquet")

Updates

Data is updated daily. Last update: 2026-05-30.

License

MIT