AgenticFinLab/PortBench-Market

PortBench Market Base Dataset

Dataset Description

A ten-year (Jan 2015–Dec 2025) daily financial dataset covering 183 instruments across six heterogeneous asset classes, designed for multi-asset portfolio management research and LLM evaluation.

Asset Coverage

Asset Class	Instruments	Data Fields	Sources
Equities	126	OHLCV + return	Yahoo Finance (ETFs: broad market, sector, factor, international)
Bonds	16	Close + return (ETFs); yield levels (FRED)	Yahoo Finance, FRED
Commodities	16	OHLCV + return	Yahoo Finance, Kaggle (spot prices)
Real Estate	10	OHLCV + return (REITs); housing indices (FRED)	Yahoo Finance, FRED
Cryptocurrency	12	OHLCV + return	Yahoo Finance, Kaggle (market cap, supply)
Cash	4	Close + return (money market ETFs); macro indicators (FRED)	Yahoo Finance, FRED

Additional Features

• Macroeconomic indicators (FRED): Fed funds rate, CPI, unemployment, GDP, yield curve spreads, breakeven inflation, credit spreads, mortgage rates, VIX, and 40+ additional series
• News text: monthly aggregated financial news for equities and cryptocurrency (JSON-encoded columns)
• Cross-asset correlation structures: intra-class and inter-class correlation matrices available as companion artifacts

Dataset Structure

Single CSV file with 4,017 rows (trading days) × 1,087 columns.

Column naming convention: {asset_class}_{ticker}_{field}

date,equities_ACWI_open,equities_ACWI_high,...,bonds_FRED_DGS10,...,cash_FRED_FEDFUNDS,...
2015-01-02,52.34,52.45,...,2.17,...,0.11,...
...
2025-12-31,...

Fields per instrument (where applicable): open, high, low, close, volume, return

Temporal Coverage

• Full range: 2015-01-02 to 2025-12-31
• Frequency: Daily (trading days)
• Three stress windows highlighted: 2015 China Shock (Aug 2015–Feb 2016), 2020 COVID Crash (Feb–May 2020), 2022 Crypto Collapse (May–Dec 2022)

Intended Use

This dataset serves as the foundation for PortBench's dual-layer evaluation. It supports:

• Static QA generation for financial reasoning tasks
• Dynamic five-stage pipeline evaluation with realistic market replay
• Stress testing under historical crisis regimes
• Cross-asset correlation analysis and portfolio optimization research

Point-in-Time (PiT) Constraint

All features strictly respect temporal integrity — no look-ahead bias. Any derived feature uses only information available at or before the corresponding date.

Citation

@article{zhao2026portbench,
  title={PortBench: A Correlation-Aware, Full-Pipeline Benchmark for LLM-Driven Portfolio Management},
  author={Zhao, Yuxuan and Chen, Sijia and Su, Ningxin},
  journal={arXiv preprint arXiv:2605.27887},
  year={2026}
}