Hugging Face's logo

Datasets:
OpenSourceAP
/
CrossSection

Modalities:
Document
Tabular
Text
Formats:
parquet
Languages:
English
Size:
1M - 10M
Tags:
finance
asset-pricing
stock-returns
anomalies
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
CrossSection / README.md
Andrew Y. Chen
README: Paper section; remove arXiv links; move branch details to Releases
ad1dfba
preview code
|
Raw History Blame Contribute Delete
25.2 kB
metadata 
pretty_name: Open Source Cross-Sectional Asset Pricing
license: cc-by-4.0
language:
  - en
tags:
  - finance
  - asset-pricing
  - stock-returns
  - anomalies
size_categories:
  - 100M<n<1B
configs:
  - config_name: long_short_returns
    data_files: Portfolios/Full Sets OP/PredictorLSretWide.parquet
    default: true
  - config_name: predictor_portfolios
    data_files: Portfolios/Full Sets OP/PredictorPortsFull.parquet
  - config_name: placebo_portfolios
    data_files: Portfolios/Full Sets OP/PlaceboPortsFull.parquet
  - config_name: signal_doc
    data_files: SignalDoc.parquet
  - config_name: predictors
    data_files: Firm Level Characteristics/Predictors.parquet

Open Source Cross-Sectional Asset Pricing (data releases)

Stock-level signals (firm characteristics) and portfolio returns from Chen and Zimmermann (2022), produced by the code at github.com/OpenSourceAP/CrossSection. The official data page and FAQ are at openassetpricing.com, which includes a browser for signal descriptions.

All past releases are available. This page shows the latest, 2026.10. See Releases for older ones, including 2021.03, which matches the published paper.

Paper

The paper for this dataset is Chen and Zimmermann (2022), "Open Source Cross-Sectional Asset Pricing," Critical Finance Review 11(2): 207–264. The data are licensed under CC BY 4.0. If you use them, please cite the paper (BibTeX).

Quick download (latest release, 2026.10)

Core predictor portfolios and documentation

Click a link to download.

Data File Size
Signal documentation SignalDoc.csv 0.2 MB
Long-short returns, original papers' construction (wide) PredictorLSretWide.parquet 1.4 MB
All portfolio returns, original papers' construction (long) PredictorPortsFull.parquet 18.7 MB
Value-weighted decile portfolio returns (long) PredictorAltPorts_DecilesVW.parquet 25.7 MB
Daily value-weighted portfolio returns, original papers' construction (long) PredictorVW.parquet 252 MB
Release notes Release Notes 2026.10.md 12 KB

Or use a parquet reader, e.g.:

import pandas as pd

ports = pd.read_parquet(
    "hf://datasets/OpenSourceAP/CrossSection@Data-Release-2026.10/Portfolios/Full Sets OP/PredictorPortsFull.parquet"
)

We offer many other portfolio implementations too. See the File Directory.

Accessing Specific Predictor Signals (Firm Level Characteristics)

For most predictors, you can quickly download them via the columns argument in pd.read_parquet().

# only the columns you list are downloaded (here 3 of the 211 in the 1.9 GB file)
bm = pd.read_parquet(
    "hf://datasets/OpenSourceAP/CrossSection@Data-Release-2026.10/Firm Level Characteristics/Predictors.parquet",
    columns=["permno", "yyyymm", "BM"],
)

Check SignalDoc.csv for the column name definitions, including the sign of each signal (e.g. higher BM implies higher mean return).

You cannot access the Predictors Price, Size, STreversal this way. To access these signals, you need a WRDS account.

The Full Panel of Firm Level Predictors (signals)

If you have a WRDS account, you can download the full dataset of predictor signals with one call using the openassetpricing (Python) or OpenSourceAP.DownloadR (R) package.

Both examples return signed signals by default: each predictor is multiplied by its Sign so that higher values imply higher expected returns. The Predictors.parquet example above returns unsigned signals.

Python:

# pip install openassetpricing
import openassetpricing as oap

openap = oap.OpenAP()  # latest release
signals = openap.dl_all_signals("pandas")  # or "polars", defaults to signed signals

R:

# install.packages("OpenSourceAP.DownloadR")
library(OpenSourceAP.DownloadR)

openap <- OpenAP$new()  # latest release
signals <- openap$dl_all_signals() # defaults to signed signals

Package documentation:

Placebos

Placebos are 114 signals whose original papers did not show clear, statistically significant predictability in long-short portfolios. This includes 14 "not-predictors," which were clearly insignificant, and 100 "indirect signals," which had only suggestive evidence or were modifications of another signal. Most come from Hou, Xue, and Zhang (2020), who count them among their 452 "anomalies." Chen and Zimmermann (2022) include them so that the dataset nests Hou, Xue, and Zhang, and helps reconcile the disparate replication rates (Hou, Xue, and Zhang's "failed replications" are mostly tests that were never shown to be predictive previously).

The placebos may still be useful. Many of them seem to have predictive power and true statistical predictability is plentiful, historically (Chen 2026, "Most Claimed Statistical Findings in Cross-Sectional Return Predictability Are Likely True," JF Insights; Chen and Dim 2023, "High-Throughput Asset Pricing").

Firm-level placebo signals are in Placebos.parquet, and their portfolio returns are in PlaceboPortsFull.parquet.

BidAskTAQ is not included. To build it with your own TAQ, ISSM and CRSP data, run the TAQ prep scripts and then BidAskTAQ.py from the CrossSection repo.

File Directory (2026.10)

Click a file or folder to open its page on the hub. Then click "Download" to download.

Older releases may have slightly different file structure.

Data Format

Firm-level signals (Firm Level Characteristics/)

  • One row per stock-month. permno is the CRSP stock identifier, yyyymm is the month as an integer (e.g. 198601), and every other column is one signal, named by its acronym in SignalDoc.csv.
  • Timing: signals are available and relevant for trading (based on the original paper) at the end of the month yyyymm. Use signals with yyyymm = 202511 to predict the return between end of November 2025 and some future date (e.g. end of December 2025).
  • Note: Many of the original papers imply that signals are valid long after the fiscal year end. For example, Fama-French 1992 assume that book equity is available 6 months after the fiscal year end, and is relevant for trading for another 12 months after that.

Monthly portfolio returns (Portfolios/)

  • Long format (PredictorPortsFull, PlaceboPortsFull and the Full Sets Alt files): one row per signal, portfolio and month, with columns
    • signalname: the signal's acronym, as in SignalDoc.csv
    • port: text, "01"–"10" for quantile portfolios and "LS" for long-short
    • date: the return month, dated the last trading day of the month
    • ret: return, in percent
    • signallag: the average lagged signal of the stocks in the portfolio
    • Nlong, Nshort: the number of stocks in the long and short legs (for quantile portfolios, Nlong is the number of stocks and Nshort is 0)
  • Wide format (PredictorLSretWide): long-short returns only, one row per month, with date and one column per predictor.
  • The return for date runs from the closing price on the last trading day of the preceding month to the closing price on the last trading day of that month.
  • Long-short portfolios are signed using the original papers' results, so the long leg is the one the original paper found to have higher returns.

Daily portfolio returns (DailyPortfolios/)

  • Long format with signalname, port, date (the trading day) and ret (in percent).

Releases

Each data release is its own branch of this repo. The latest release is on branch Data-Release-2026.10, and the main branch holds a copy of it (plus this README), so downloads that don't name a branch (no revision= in Python) get the latest data.

For another release, replace Data-Release-2026.10 in the link with that release's branch from the table below.

Release Branch (revision=) Notes
2026.10 "Data-Release-2026.10" 212 predictors, 114 placebos; data through 2025-12; CRSP v2 (CIZ); parquet only. See Release Notes 2026.10.md
2025.10 "2025.10" 212 predictors, 114 placebos; data through 2024-12; code translated from Stata to Python (v2.0.0)
2024.10 "2024.10" 212 predictors, 114 placebos; patch of 2024.08 fixing look-ahead bias in AnnouncementReturn
2024.08 "2024.08" 212 predictors, 114 placebos; data through 2023-12; OptionMetrics signals end 2022-12
2023.08 "2023.08" 212 predictors, 114 placebos; data through 2022-12; 5 new predictors; BM follows Stattman (1980)
2022.04 "2022.04" 207 predictors, 114 placebos; SignalDoc corrections; no daily portfolios or Results/
2022.03 "2022.03" 207 predictors, 114 placebos; data through 2021-12; SignalDoc.csv replaces the xlsx; FF93-style portfolios added
2021.04 "2021.04" 205 predictors, 114 placebos; adds value-weighted and daily portfolios; docs in SignalDocumentation.xlsx
2021.03 "2021.03" 205 predictors, 114 placebos; code v1.0.0, matches the CFR publication; portfolios in Full_Sets_Main/ and Full_Sets_Extended/; signed_predictors_dl_wide lacks KZ and ZScore
2020.07 "2020.07" 210 base and 104 additional signals; adds trading-cost portfolios (PortCosts/); original layout
2020.05 "2020.05" 210 base and 104 additional signals; first release; original layout (signalbase, portbase, ...)

Counts include Price, Size, STreversal and BidAskTAQ, which are not in the firm-level files.

Older releases were originally distributed as csv/zip files on Google Drive. On this hub, every csv (zipped or not) is stored as parquet at the same path. The other files (pdf, tex, xlsx, docx) are kept as they were. The layout of each release follows its original Drive folder, so file paths can differ between releases. To see what a release contains, list its files:

from huggingface_hub import list_repo_files

list_repo_files("OpenSourceAP/CrossSection", repo_type="dataset", revision="Data-Release-2026.10")

Loading a release

Python, one file at a time (cached locally after the first download):

import pandas as pd
from huggingface_hub import hf_hub_download

release = "Data-Release-2026.10"  # any branch in the table above

path = hf_hub_download(
    repo_id="OpenSourceAP/CrossSection",
    repo_type="dataset",
    filename="Firm Level Characteristics/signed_predictors_dl_wide.parquet",
    revision=release,
)
signals = pd.read_parquet(path)

One line per release, for the main portfolio returns:

from huggingface_hub import hf_hub_download

f = "Portfolios/Full Sets OP/PredictorPortsFull.parquet"
kw = dict(repo_id="OpenSourceAP/CrossSection", repo_type="dataset", filename=f)

hf_hub_download(**kw, revision="Data-Release-2026.10")
hf_hub_download(**kw, revision="2025.10")
hf_hub_download(**kw, revision="2024.10")
hf_hub_download(**kw, revision="2024.08")
hf_hub_download(**kw, revision="2023.08")
hf_hub_download(**kw, revision="2022.04")
hf_hub_download(**kw, revision="2022.03")
hf_hub_download(**kw, revision="2021.04")

(2021.03 keeps this file under Portfolios/Full_Sets_Main/. The 2020 releases use the original layout, so call list_repo_files first.)

With 🤗 datasets:

from datasets import load_dataset

ports = load_dataset(
    "OpenSourceAP/CrossSection",
    data_files="Portfolios/Full Sets OP/PredictorPortsFull.parquet",
    revision="Data-Release-2026.10",
    split="train",
)

pandas can read straight from the hub (via huggingface_hub). Put the release after @:

import pandas as pd

ports = pd.read_parquet(
    "hf://datasets/OpenSourceAP/CrossSection@Data-Release-2026.10/Portfolios/Full Sets OP/PredictorPortsFull.parquet"
)

# only the columns you list are downloaded (here 3 of the 211 in the 1.9 GB file)
bm = pd.read_parquet(
    "hf://datasets/OpenSourceAP/CrossSection@Data-Release-2026.10/Firm Level Characteristics/Predictors.parquet",
    columns=["permno", "yyyymm", "BM"],
)

columns= saves downloading only with hf:// paths. With an https:// link, pandas downloads the whole file first.

R (download, then read with arrow):

release <- "Data-Release-2026.10"
url <- paste0(
  "https://huggingface.co/datasets/OpenSourceAP/CrossSection/resolve/", release,
  "/Portfolios/Full%20Sets%20OP/PredictorPortsFull.parquet"
)
download.file(url, f <- tempfile(fileext = ".parquet"), mode = "wb")
ports <- arrow::read_parquet(f)

License

The data are licensed under CC BY 4.0. Please cite the paper below.

Citation 

@article{ChenZimmermann2022,
  title   = {Open Source Cross Sectional Asset Pricing},
  author  = {Chen, Andrew Y. and Zimmermann, Tom},
  journal = {Critical Finance Review},
  year    = {2022},
  volume  = {11},
  number  = {2},
  pages   = {207--264}
}