Skip to article content

Revised 23 October 2026 · v1.0.0 · what changed

Code: github.com/herrrickshaw/piotroski-liquidity-research

Watch the walkthrough

Watch on YouTube: youtu.be/m8uKeoLRaBo

Why it is useful

• Who benefits: students and independent researchers who want a worked, honest example of a point-in-time factor study, including the data bugs found along the way.
• What it replaces: guessing whether an F-score screen survives size and liquidity controls; the repo ships the scripts, the paper and the result tables.
• Evidence from the run: the US scripts (completeness audit, daily Darvas x F-score scan, weight-preset sweep, cost-versus-edge) all ran end to end on 23 October 2026 and regenerated their result files in under a minute each.
• Evidence from the run: in the regenerated daily signal file (16,333 signals, 1,802 tickers) the median 12-month forward return rose from -2.8% for the lowest-turnover tier to +8.9% for the highest, the same direction as the committed file (-3.1% to +8.2%). Almost all of those signals fall in 2023 to 2026, so it is a short, regime-dependent window.

What it is for

The repo studies one question: is the Piotroski F-score premium conditioned by firm size or by liquidity? It double-sorts liquidity within size, uses point-in-time market cap and SEC EDGAR filing dates, keeps delisted names, and reports medians rather than means.

Its headline table says the premium is largest in large but illiquid firms and comes from weak-F names falling, not strong-F names rising. It then estimates execution costs and puts a capacity ceiling near $250,000.

The README opens with a warning that its own figures are being regenerated after data defects were found, so treat every number as provisional. That candour is part of what makes it worth reading.

The scripts need an EDGAR filings table and a US price panel that the repo does not ship. For this post I re-ran the US scripts on a US price panel (last session 21 October 2026) and an EDGAR filings table (filed up to 20 October 2026), with the hard-coded file paths pointed at those copies in a scratch clone. The EDGAR table is dense only from fiscal 2021, so the re-run covers roughly 2023 to 2026, not the repo's 2016 to 2025. I did not use the India scripts or file, which derive from screener.in data.

Who it is for

• Students learning point-in-time backtesting, survivorship handling and why medians beat means
• Researchers who want a template for liquidity-conditioned factor tests
• Anyone sceptical of factor claims who wants to see the validation failures documented

What you need

Get the code and an AI assistant

You do not have to be a programmer. Get the repository, give it to an AI assistant, and ask in plain language. Any of these routes works:

git clone https://github.com/herrrickshaw/piotroski-liquidity-research.git
cd piotroski-liquidity-research
claude      # or: gemini

Install

1. Clone the repo and make a virtual environment.

   git clone https://github.com/herrrickshaw/piotroski-liquidity-research.git
   cd piotroski-liquidity-research
   python3 -m venv .venv && source .venv/bin/activate
   pip install pandas numpy duckdb
   # then edit the FUND and PX constants at the top of each US script

Quick start: commands and what you should see

Check the sample-size claim before trusting anything (run this first, as the README says).

python3 data_completeness_audit.py

One line: fundamentals tickers, price tickers and their overlap. On 23 Oct 2026: 6,897 | 10,110 | 6,889 (100% / 68%).

Regenerate the daily Darvas x F-score signal file and the tier table.

python3 daily_breakout_combo_us.py

About 40 seconds. Prints signal counts, a tier table and a per-year table, and rewrites reports/daily_breakout_combo_us.csv.

Run the weight-preset sweep, then the cost test that reads its output.

python3 sweep_piotroski_plus_us.py && python3 cost_vs_edge.py

About 15 seconds each. The cost test prints net edge by tier at ten portfolio sizes.

Sample outputs

Real output from a run on the date shown in the sources line; trimmed to fit.

Data completeness audit (23 Oct 2026; US price panel last session 2026-10-21)

python3 data_completeness_audit.py
=== source reconciliation ===
  fundamentals  6,897 | prices 10,110 | overlap  6,889  (100% / 68%)

The committed script comments quote 9,278 price symbols and 4,459 overlap (97%). The fundamentals table here has 6,897 tickers, but most fiscal years before 2021 are sparse, so tickers overlap does not mean years of history overlap.

Daily Darvas x F-score scan, by liquidity tier (regenerated; winsorised median minus same-tier baseline)

python3 daily_breakout_combo_us.py
  raw first-cross breakouts: 471,700
  after 10-bar cooldown: 201,749 signals, 7,960 symbols
  signals WITH point-in-time fundamentals: 16,333 (1,802 symbols, 11 years)

  tier     base med    darvas      n   darvas×F7     n  darvas×F<=3     n
  SMALL        0.1%     -2.9%  5,445       +6.6%   850       -14.9%  1076
  MID          2.9%     +3.3%  5,444       +7.9%  1695        -7.7%   426
  LARGE        6.8%     +2.1%  5,444       +2.7%  1984        -6.5%   248

By signal year: 4 in 2016, 16 in 2017, 5 in 2018, 17 in 2019, 31 in 2020, 31 in 2021, 127 in 2022, 4,059 in 2023, 5,887 in 2024, 5,800 in 2025, 356 in 2026. Effectively a three-year sample, and signals cluster in rallies.

Median 12-month forward return by tier, then by tier and F-score band (re-aggregated from the regenerated file)

groupby('tier') and pd.cut(F9,[-1,3,6,9]) on reports/daily_breakout_combo_us.csv
          n     med    mean      sd
LARGE  5444  0.0894  0.2146  0.6170
MID    5444  0.0619  0.1774  0.6074
SMALL  5445 -0.0280  0.1438  0.7874

        F0-3     F4-6     F7-9
LARGE   0.0036   0.0855   0.1018
MID    -0.0487   0.0565   0.0882
SMALL  -0.1480  -0.0452   0.0609

The committed file showed no clean F-score gradient (for example SMALL -1.2%, -3.8%, -2.1%). In this 2023 to 2026 window every tier rises from F0-3 to F7-9. Cells range from 248 to 3,149 rows; 252-day windows overlap.

Weight-preset sweep, US (regenerated; 9 rebalances, but only 4 with forward returns)

python3 sweep_piotroski_plus_us.py
  fundamentals: 15,654 ticker-years | 3,865 tickers
  scored panel: 7,548 stock-years | 2,926 symbols | 9 rebalances

  vector         mean  median  vs univ  beat univ   worst
  [universe]    34.4%   15.0%       --         --  -32.6%
  canonical     29.4%   28.1%   +14.8%    4/4       13.0%
  quality       30.1%   28.1%   +15.6%    4/4       16.0%
  turnaround    24.2%   21.1%   +9.7%    3/4       11.6%
  safety        28.4%   28.1%   +13.9%    4/4        9.4%

  tier         n  F>=70 med  F<40 med  med edge
  SMALL    2,518       1.2%    -25.4%    +26.5%
  MID      2,516       5.5%      5.4%     +0.1%
  LARGE    2,514      13.9%     19.9%     -6.0%

Not comparable with the committed sweep (which had 3,294 observations): different universe, and only four rebalance years. The universe mean (34%) far above its median (15%) points to outliers, probably unadjusted corporate actions in the price panel, so read medians only. The small-tier edge is the opposite sign to the large tier here, the pattern Piotroski's original paper predicts, per a comment in the script.

Does the small-cap edge survive costs? (regenerated cost test, 2,923 symbols; prices through 2026-10-21)

python3 cost_vs_edge.py
  tier       n   turnover/day    spread   impact per $1M
  SMALL    975          $0.2M     1.07%          47.312%
  MID      974          $5.8M     0.48%           0.648%
  LARGE    974         $75.7M     0.23%           0.026%

  $    100,000  cost   2.01%/yr  (  6.0% of ADV)  best gross +13.3% -> net  +11.3%  SURVIVES
  $    250,000  cost   3.43%/yr  ( 15.0% of ADV)  best gross +13.3% -> net   +9.9%  SURVIVES
  $    500,000  cost   5.80%/yr  ( 30.0% of ADV)  best gross +13.3% -> net   +7.5%  SURVIVES
  $  1,000,000  cost  10.53%/yr  ( 60.0% of ADV)  best gross +13.3% -> net   +2.8%  SURVIVES
  $  2,000,000  cost  19.99%/yr  (119.9% of ADV)  best gross +13.3% -> net   -6.7%  DEAD

Two things. The README's capacity ceiling of about $250k is consistent with these costs (15% of daily volume at $250k). But the script defines a 20%-of-volume ceiling (ADV_CEILING_PCT) and never applies it, so it still prints SURVIVES at $500k and $1M, where positions are 30% and 60% of daily volume. I saved a patch (not pushed). The gross edges used are constants hard-coded in the script from the earlier sweep, not recomputed.

Ways to use it

Sample prompts for your AI assistant

Copy, paste and edit the parts in [brackets].

• Read README.md and the paper in reports/ and explain, in plain English, why the author says the premium tracks illiquidity rather than size.
• Using reports/daily_breakout_combo_us.csv, compute the median and win rate of wret by tier and by year and tell me which results are most regime-dependent.
• List every data defect the README says was found, and for each say whether it is fixed and which results it could change.
• Explain why the README treats median as the headline and what overlapping 252-day windows do to the effective sample size.
• Write a checklist I could apply to my own backtest from the 'Metrics that failed validation' table.
• Suggest what extra controls (beta, book-to-market) would be needed before treating this as evidence, and what data I would need.

Limits and honest caveats

Sources: README, paper and scripts in github.com/herrrickshaw/piotroski-liquidity-research, re-run on 23 October 2026 in a throwaway virtual environment against a US price panel (last session 21 October 2026) and an SEC EDGAR filings table (filed to 20 October 2026). Underlying filings are SEC EDGAR public data; only derived aggregates shown; no vendor data.

About this article: Researched, written and edited by Umashankar Triplicane Dwarakanathan, with AI research assistance; every figure is meant to trace to the primary source cited. See the Editorial Policy for how sourcing, AI use and corrections work.