README.md

metadata

license: other
task_categories:
  - feature-extraction
  - summarization
  - tabular-to-text
  - table-to-text
  - text-retrieval
tags:
  - finance
  - investing
  - portfolio
  - factors
  - synthetic-data
  - hedge-funds
  - buffett
  - alpha

Synthetic Buffett's Alpha Dataset

This dataset contains realistic synthetic data inspired by the paper “Buffett’s Alpha” (Frazzini, Kabiller, Pedersen, 2018).

It simulates Berkshire Hathaway–style returns, leverage, financing, factor exposures, portfolio decomposition, and systematic replication performance.

All data is artificially generated using stochastic processes (GBM, GARCH-like models) and contains no proprietary or private financial data.

---

Dataset overview

• The dataset includes 100 independent synthetic companies (BuffettCo_1 … BuffettCo_100).
• Each company has 1,000 monthly observations (~83 years of synthetic data).
• All data is stored in a single unified file: buffett_alpha_synthetic.csv.
• Each row corresponds to one company × month observation.

Total size: ~100,000 rows × ~25 columns.

---

Columns included

Identification

• scenario_id — numeric ID of the simulated company
• company_name — company label (e.g., BuffettCo_17)
• date — monthly timestamp

Returns & performance

• market_excess_return
• berkshire_excess_return
• sharpe_ratio (constant across a company’s rows)
• info_ratio (constant across a company’s rows)

Leverage & financing

• leverage_ratio
• insurance_float_cost
• debt_outstanding
• float_share_of_liabilities

Factor exposures

• MKT (market beta)
• SMB (size)
• HML (value)
• UMD (momentum)
• BAB (betting against beta)
• QMJ (quality minus junk)
• Alpha

Portfolio decomposition

• public_stock_return
• private_company_return
• public_weight
• private_weight
• combined_return

Replication comparison

• actual_berkshire_return
• systematic_buffett_style_return
• tracking_error

---

Relation to the original paper

In the original Buffett’s Alpha study:

• The authors used real-world data from 1976–2017 (~500 monthly observations).
• There was only one company: Berkshire Hathaway.

In this dataset:

• There are 100 Buffett-style companies, not just one.
• Each has 1,000 months of synthetic data, much larger than the original sample.

This enables:

• Reproducing the types of analyses from the paper (Sharpe ratios, factor regressions, leverage effects).
• Exploring alternative “what if” Buffett-style histories.
• Using long time series for machine learning, econometrics, or stress testing.

---

Example usage

from datasets import load_dataset

# Load the dataset (replace with your HF username/repo)
ds = load_dataset("your-username/buffett-alpha-synthetic")

# Convert to pandas for analysis
df = ds["train"].to_pandas()  # single split
print(df.head())

# Example: compute mean excess return by company
print(df.groupby("company_name")["berkshire_excess_return"].mean().head())