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 / Release Notes 2026.10.md
Andrew Y. Chen
Update release notes for 2026.10
7386135
preview code
|
Raw History Blame Contribute Delete
12.1 kB

Documentation for Open Source Cross-Sectional Asset Pricing Data

Andrew Chen and Tom Zimmermann

Data Release: October 2026

Created with Code Version: 2.1.0

New data relative to the October 2025 release

  • Data through December 2025, with the following Exceptions:
    • Predictors based on OptionMetrics implied vol (CPVolSpread, dCPVolSpread, dVolCall, dVolPut, RIVolSpread, skew1, SmileSlope) are still taken from our August 2023 release and end in December 2022
    • Predictors based on the Gompers, Ishii, Metrick governance index (Activism1, Activism2, and Governance) end in 2007 as that data is no longer updated
    • PatentsRD uses NBER patent data that ended in 2006.
    • ProbInformedTrading relies on the Duarte et al. data which ends in 2012. See issue #171
  • CRSP data now comes from CRSP v2 (CIZ format) instead of the legacy SIZ format, which WRDS is retiring
    • Differences between v1 and v2 outputs, and how we handled them, are documented in Docs/CRSP-v2-Migration.md
    • The switch moves portfolio returns little. The R-sq from regressing this release's long-short returns on the previous release leads to an R-sq > 98% for 75% of predictors.

Data format update

  • Files are now parquet instead of zipped csvs, and each set of signals or portfolios is one file. See Directory below
  • The redundant individual signals and individual portfolios are no longer posted.
  • These changes were made as part of a migration from Google Drive to Hugging Face

Signals with Non-Trivial Revisions

Non-trivial revisions means R-sq < 95% from regressing this vintage's monthly long-short returns on the previous vintage's (full sample). 26 of 212 predictors qualify. R-sq in parentheses.

  • Traceable to issue fixes
    • PayoutYield (48.6), NetPayoutYield (78.5): fix preferred stock handling (pstkrv is a stock, not a flow) and improve replication
    • CompositeDebtIssuance (79.4): drop infinite values from log(x/0), and match debt 60 months earlier with a proper calendar lag instead of a positional shift
    • Other signals using calendar lags: standardized so missing months are not filled with the wrong observation, and months inserted to compute lags are no longer kept in the output. Signals that moved: CitationsRD (59.1), MomOffSeason11YrPlus (88.4), MomOffSeason16YrPlus (91.2), MomOffSeason06YrPlus (92.4)
    • IntanEP (78.8), IntanSP (80.5), IntanCFP (82.8): follow the original paper by taking logs of all ratios and using income before extraordinary items (ib) instead of net income throughout
    • High52: split-adjust prices, and include the current month in the 52-week high, following George and Hwang (2004)
    • Placebos: about 10 placebo scripts now follow the original Stata code more closely
  • Traceable to the CRSP v2 migration
    • SIC codes: from about 2008 the legacy CRSP data left many new listings unclassified (SIC 9999). Apparently CRSP v2 did not classify these listings either as of July 2024 (Schwarz, Walter and Weiss find almost no SIC differences between v1 and v2 in data from that time). As of October 2026, CRSP v2 assigns them codes, which changes industry membership. Affects EarnSupBig (77.0), Herf (77.9), HerfAsset (82.8), HerfBE (84.0), MS (89.4), OrgCap (89.6), ChInvIA (89.7), IndMom (91.7), sinAlgo (92.7), and IndRetBig (94.7)
    • ExchSwitch (81.7): CRSP v2 assigns exchange codes to about 21,500 stock-months that the legacy data left unclassified
  • Traceable to data revisions
    • ChangeInRecommendation (93.6), UpRecomm (93.7), ConsRecomm (93.9), DownRecomm (94.5): IBES has removed or revised some historical analyst recommendations. Values are unchanged where both vintages have data
  • Traceable to floating-point precision
    • Investment (93.0): firm-months with zero capital expenditures throughout the 36-month averaging window are now mostly missing (0/0), matching the Stata original. The previous release recorded about 8,800 of them as zero because of floating-point residue in the rolling mean, which still affects about 200 firm-months in this release
  • Source of revision unclear
    • None

Fixes that do not change returns:

  • 13F split adjustments are now done in one isolated block early in the code, to make clear there is no look-ahead bias. See issue #155
  • dCPVolSpread: fixed the sign of the Return column in SignalDoc.csv. See issue #206
  • Illiquidity.csv columns are now in the standard order (permno, yyyymm, Illiquidity)
  • SignalDoc browser: fixed broken code links for signals made by multi-signal scripts. See issue #204

Directory

All data files are parquet, readable with pandas.read_parquet or polars.read_parquet in Python and arrow::read_parquet in R. Each set of signals or portfolios is a single file. There are no zip files and no per-signal csvs. The same files are on Hugging Face at OpenSourceAP/CrossSection, branch Data-Release-2026.10.

  • SignalDoc.csv and SignalDoc.parquet
    • Describes each signal and contains hand-collected statistics. The two files have the same content
  • The SignalDoc browser lets you browse summary stats for all reproduced signals
  • storage_checks.txt
    • Automated checks run on this folder before release, including the summary stats above
  • Firm Level Characteristics/
    • Note: downloadable predictors do not include size, price, or short-term reversal. To make them, run Signals/pyCode/crsp_only.py in our code, which downloads CRSP and builds Size, Price, and STreversal. We also do not include bid-ask spreads from TAQ, but that was not shown to predict returns anyway.
    • Predictors.parquet
      • All downloadable predictors in wide format. Columns: permno, yyyymm, then one column per predictor, named by its acronym in SignalDoc.csv. One row per permno-month that has at least one predictor; missing values are null
    • signed_predictors_dl_wide.parquet
      • Same as Predictors.parquet, but each predictor is multiplied by its Sign in SignalDoc.csv, so a higher signal implies higher mean returns based on the original papers
    • Placebos.parquet
      • Same layout as Predictors.parquet, but for characteristics that were not predictive or were indirect signals based on the original results. See SignalDoc.csv for details
  • Portfolios/
    • Unless noted, portfolio files are long, with columns: signalname, port, date, ret, signallag, Nlong, Nshort. port is the portfolio name (01, 02, ..., LS for long-short), ret is the monthly return in percent during the month indicated by date, and signallag is the portfolio's average lagged signal
    • Full Sets OP/
      • Full sets of portfolios following the original papers (OP)
      • PredictorPortsFull.parquet
        • Portfolios formed on each predictor (e.g. 5 portfolios formed by sorting on momentum), plus the long-short portfolio
        • Includes what some people call "test asset returns." Can also be used to study monotonicity
        • Implementations based on the original papers. See SignalDoc.csv
      • PredictorLSretWide.parquet
        • Long-short returns only, in wide format. Columns: date, then one column per predictor
        • These are performance of trading strategies based on cross-sectional predictors, or, if you like, realized factor premiums
      • PredictorSummary.xlsx
        • Summary stats for all predictor portfolios
      • PlaceboPortsFull.parquet and PlaceboSummary.xlsx
        • Portfolios and summaries based on not-predictors and indirect signals
    • Full Sets Alt/
      • Full sets of portfolios with alternative implementations
      • PredictorAltPorts_Deciles.parquet
        • Like PredictorPortsFull, but only continuous predictors, and implemented as deciles. Stock weights follow OP
      • PredictorAltPorts_DecilesEW.parquet and PredictorAltPorts_DecilesVW.parquet
        • Like PredictorAltPorts_Deciles, but stock weights are all EW or all VW
      • PredictorAltPorts_Quintiles.parquet, PredictorAltPorts_QuintilesEW.parquet, and PredictorAltPorts_QuintilesVW.parquet
        • Same as the decile files but with quintile sorts
      • PredictorAltPorts_FF93style.parquet
        • Like PredictorPortsFull, but uses Fama-French 1993 style 2x3 sorts
      • PredictorAltPorts_HoldPer_1.parquet, _3, _6, and _12
        • Like PredictorPortsFull, but uses "rebalancing periods" of 1, 3, 6, and 12 months instead of the original rebalancing periods
        • These rebalancing periods should really be called signal updating periods, since value-weighting or equal-weighting is always enforced monthly, see footnote in the paper
      • PredictorAltPorts_LiqScreen_ME_gt_NYSE20pct.parquet, _NYSEonly, _Price_gt_5, and _VWforce
        • Like PredictorPortsFull, but with various liquidity adjustments. Please see paper
      • PlaceboPortsFull.parquet
        • A copy of the file in Full Sets OP/
  • DailyPortfolios/
    • Daily portfolio returns. Aggregates up to the monthly strategies (almost). Only contains returns (%, daily), please see monthly data for supporting statistics like # of stocks
    • Each file is long, with columns: signalname, port, date, ret
    • Predictor.parquet
      • Following original papers
    • Other implementations match the monthly ones (see Portfolios / Full Sets Alt)
      • PredictorVW.parquet
      • CtsPredictorDecile.parquet
      • CtsPredictorDecileVW.parquet
      • CtsPredictorQuintile.parquet
      • CtsPredictorQuintileVW.parquet
    • DailyPortSummary.xlsx
      • Some summary stats for the daily portfolios, since we don't provide results in the paper
      • Sumstats sheet should be self explanatory. Note number of signals in each predictor port varies because the original paper vary in the number of portfolios they form
      • Timingcheck: regressions of monthly returns on daily returns aggregated to monthly, done by groups of (signalname, portfolio)
      • Further details, see daily portfolio construction R script
  • Results/
    • Figures (pdf) and tables (tex) for the paper, rebuilt with this release's data