README.md

metadata

license: mit
language:
  - en
size_categories:
  - 1GB < n < 10GB
tags:
  - finance
  - portfolio-management
  - multi-asset
  - benchmark
  - LLM
  - correlation

PortBench-RawData

This repository contains the raw collected data and preprocessed asset files for PortBench.

The data spans 2015–2025 across six heterogeneous asset classes: Equities, Bonds, Commodities, Real Estate, Cryptocurrency, and Cash.

---

Repository Structure

PortBench-RawData/
├── raw_data/                          # Raw collected data (~4.6 GB)
│   ├── fred/                          # FRED macroeconomic indicators (60 series)
│   │   ├── bonds/                     # Yield curve, credit spreads, TIPS
│   │   ├── cash/                      # Fed funds rate, CPI, GDP, employment
│   │   ├── commodities/               # Oil, gold, agriculture spot indices
│   │   └── real_estate/               # Case-Shiller, HPI, REIT indices
│   ├── kaggle/                        # Kaggle supplementary data (~4 GB)
│   │   ├── commodities/               # Commodity futures and spot prices
│   │   ├── cryptocurrency/            # Crypto OHLCV data
│   │   ├── equities/                  # Stock data with news text
│   │   └── real_estate/               # REIT and property data
│   ├── sec/                           # SEC EDGAR filings
│   │   └── equities/                  # 10-K, 10-Q filings for US equities
│   ├── yahoo/                         # Yahoo Finance price data
│   │   ├── bonds/                     # Bond ETF prices and yields
│   │   ├── cash/                      # Money market and treasury ETF data
│   │   ├── commodities/               # Commodity ETF data
│   │   ├── cryptocurrency/            # Crypto ETF and trust data
│   │   ├── equities/                  # 72 tickers (broad-market, sector, factor ETFs)
│   │   └── real_estate/               # REIT ETF data
│   └── metadata.json                  # Dataset metadata summary
│
└── processed/                         # Preprocessed asset data
    ├── equities.csv                   # 126 equity tickers, aligned daily
    ├── bonds.csv                      # 15 bond series, aligned daily
    ├── commodities.csv                # 16 commodity series, aligned daily
    ├── real_estate.csv                # 10 real estate series, aligned daily
    ├── cryptocurrency.csv             # 12 cryptocurrency series, aligned daily
    ├── cash.csv                       # 4 cash equivalent series, aligned daily
    ├── correlation_matrix.csv         # 183×183 Pearson correlation matrix
    ├── asset_class_map.json           # Ticker-to-asset-class mapping
    └── time_ranges.json               # Per-ticker date coverage ranges

---

Data Sources

Source	Coverage	Content	Tickers/Series
Yahoo Finance	2015–2025	Daily OHLCV, adjusted close, volume for ETFs and stocks across all 6 asset classes	72 tickers
FRED	2015–2025	Macroeconomic indicators: yield curve (DGS1–DGS30), TIPS real yields (DFII5/10/30), breakeven inflation (T5YIE/T10YIE), Fed funds rate (DFF/FEDFUNDS), CPI (CPIAUCSL/CPILFESL), GDP, employment (PAYEMS), housing (CSUSHPINSA), VIX	60 series
Kaggle	2015–2025	Supplementary cryptocurrency OHLCV, commodity futures, equities with news sentiment, real estate data	~45 datasets
SEC EDGAR	2015–2025	10-K and 10-Q filings for US equities, parsed text	Equity filings

---

Coverage by Asset Class

Asset Class	Raw Tickers	Processed	Role in Portfolio
Equities	127	126	Return engine; diversified via sector/factor ETFs (SPY, QQQ, XLE, XLF, VXUS, etc.)
Bonds	15	15	Fixed-income hedging; Treasury, corporate, high-yield (TLT, IEF, HYG, LQD, etc.)
Commodities	16	16	Inflation hedge; gold, oil, natural gas, agriculture (GLD, USO, UNG, DBC, etc.)
Real Estate	10	10	Diversification; REIT sector ETFs (VNQ, IYR, SCHH, etc.)
Cryptocurrency	12	12	High-risk allocation; major + mid-cap (BTC, ETH, SOL, DOGE, etc.)
Cash	4	4	Capital preservation; money market, short-term Treasuries (BIL, SGOV, SHV)

Cross-asset correlations exhibit the key structural property exploited by PortBench's dual-layer scoring: intra-class correlations are strongly positive (0.4–0.6+), while inter-class correlations are near-zero or negative, meaning true diversification requires cross-class allocation, not just many tickers within one class.

---

Preprocessing Details

• Calendar alignment: All series aligned to a common business-day calendar; gaps ≤5 days forward-filled; longer gaps retained as NaN for pairwise-complete correlation estimation.
• Market regime labels: Each asset class labeled as bull/bear/sideways/crisis using MA crossover (50/200-day) + 15% max-drawdown crisis threshold.
• Data splits: Train (2015–2022), Validation (2023–2024), Test (2025), with year-end boundaries.
• Correlation matrix: 183×183 Pearson correlation matrix computed from daily simple returns over the full training period using pairwise-complete observations; frozen and not re-estimated.

---

Related Datasets

This repository contains the raw data foundation. See also:

• PortBench-Market — The processed market base dataset (portbench.csv) with all six asset classes merged at daily frequency, plus visualization figures.
• PortBench-QA — 6,269 question-answer pairs across 7 templates (T1–T7) probing correlation-based financial reasoning, with train/val/test splits.

Both are part of the PortBench collection.

---

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}
}

---

License

This dataset is released under the MIT License. Data sourced from Yahoo Finance, FRED, Kaggle, and SEC EDGAR is subject to their respective terms of service.