Phase 0 — the historical test
The Financier 2026-09-18, H3 added 2026-09-19, H3 sourcing verified 2026-09-26, H3 parent-resolution tested 2026-10-03
Read this before any number below
These results are an upper bound on what this method would have done, not an estimate of it. Four things push the numbers the same way — up — and none of them is fully fixable from free data.
Survivorship bias. Daily prices come from a free source that drops companies once they delist. A company that went to zero is usually absent from the data rather than present and down 100%. Both hypotheses buy small and mid-size companies, which is exactly where this bites hardest. Treat every return below as flattered by an unknown amount, and treat the gap between the two hypotheses as more trustworthy than either one's level.
Lookahead bias. Guarded against, deliberately and at a cost to the results. Every fundamental and every insider filing enters the simulation on its filing date, never on the date of the period it describes or the trade it reports. The XBRL frames API would have made the market-cap screen far easier and was rejected for exactly this reason: it serves the most recently filed value for a historical period, so a 2017 number restated in 2020 would have arrived three years early.
Transaction costs. Assumed at 0.2% per side and charged on the portfolio fraction actually traded each month. Every table below shows the result before and after. No slippage, no market impact, and no bid-ask spread beyond that 0.2% — all three are worse in reality for small companies than for large ones.
Overfitting. The rules were fixed in charters/financier.md before the data was downloaded and were not tuned. No parameter below was chosen by trying alternatives and keeping the best. That is the single most important line in this report, because it is the one that cannot be verified by looking at the results — only by looking at the git history, where the charter's commit precedes the data.
One more, which the charter does not name and this run ran into anyway: a paper backtest has no fear. It holds through every drawdown below without flinching, because it is arithmetic. A person holding the same positions would have had to decide, in the month it was down, to keep holding.
H1 — Opportunistic cluster insider buying
The rule, fixed before the data was looked at: within any 30-day window, two or more distinct officers or directors of the same company make open-market purchases (Form 4, code P) totalling at least $100,000, and none is a routine buyer (same insider bought in the same calendar month in at least 2 of the prior 3 years). Enter at the next month's open, equal weight, hold 6 months, rebalance monthly. Basis: Cohen, Malloy & Pomorski, "Decoding Inside Information", Journal of Finance 2012.
Verdict: FAIL
The decision rule was written before these results existed and is applied as written: a hypothesis passes only if it beats VTI after assumed costs and tax drag in both windows, with a maximum drawdown no more than 1.5× VTI's.
- Full window (2014-01 to 2025-12): After costs and tax drag it does not beat the benchmark (6.67% against 12.88%), and its maximum drawdown of -35.12% is within the 1.5× limit (35.12% against a limit of 37.23%).
- 2020–2025 (2020-01 to 2025-12): After costs and tax drag it does not beat the benchmark (6.06% against 14.37%), and its maximum drawdown of -35.12% is within the 1.5× limit (35.12% against a limit of 37.23%).
This is a fail, and it is a finished result. It is not softened into a partial pass, and it does not earn a forward paper position. Per the charter, the Financier moves to the next hypothesis in the queue (spin-offs; post-earnings-announcement drift).
Full window, 2014-01 to 2025-12
| Measure | Strategy | VTI | Label |
|---|---|---|---|
| Annualised return, before costs | 11.14% | 12.88% | sourced |
| Annualised return, after 0.2%/side costs | 10.26% | 12.88% | assumed cost |
| Annualised return, after costs and tax drag | 6.67% | 12.88% | assumed 35% short-term |
| Annualised volatility | 22.28% | 14.93% | sourced |
| Maximum drawdown | -35.12% | -24.82% | sourced |
| Drawdown peak → trough | 2021-06 → 2022-09 | 2021-12 → 2022-09 | sourced |
| Hit rate of individual positions | 55.01% | — | sourced |
| Completed positions | 5777 | — | sourced |
| Average positions held | 285 | 1 | sourced |
| Annual turnover, one-way | 200.06% | 0% | sourced |
2020 to 2025 alone, so a recent-regime failure is visible
| Measure | Strategy | VTI | Label |
|---|---|---|---|
| Annualised return, before costs | 10.21% | 14.37% | sourced |
| Annualised return, after 0.2%/side costs | 9.32% | 14.37% | assumed cost |
| Annualised return, after costs and tax drag | 6.06% | 14.37% | assumed 35% short-term |
| Annualised volatility | 26.51% | 17.67% | sourced |
| Maximum drawdown | -35.12% | -24.82% | sourced |
| Drawdown peak → trough | 2021-06 → 2022-09 | 2021-12 → 2022-09 | sourced |
| Hit rate of individual positions | 51.95% | — | sourced |
| Completed positions | 3765 | — | sourced |
| Average positions held | 354 | 1 | sourced |
| Annual turnover, one-way | 203.25% | 0% | sourced |
How this was built
- Signal source: the SEC's quarterly Form 3/4/5 structured datasets,
https://www.sec.gov/dera/data/form-345(sourced). - 14254 qualifying cluster signals across 4897 distinct companies (sourced).
- 2087 of those companies (43%) had no usable price history and were dropped. Most are delistings and acquisitions, and dropping them is the survivorship bias described above, made concrete.
- The 30-day cluster window runs on transaction date; the position is entered at the open of the month after the latest filing date in the window, because a Form 4 is public when it is filed and not when the trade happened (assumed implementation of a rule that does not specify it, chosen against the result).
- The routine-buyer exclusion counts only years strictly before the window, so the future never decides who was a routine buyer.
H2 — Momentum with a quality filter
The rule, fixed before the data was looked at: each month rank U.S. stocks above $500M market cap by 12-month return skipping the most recent month; take the top decile; drop any name with negative trailing-12-month operating income; equal weight, hold 1 month. Basis: Jegadeesh & Titman, Journal of Finance 1993, with the quality overlay per Asness, Frazzini & Pedersen, "Quality Minus Junk", Review of Accounting Studies 2019.
Verdict: FAIL
The decision rule was written before these results existed and is applied as written: a hypothesis passes only if it beats VTI after assumed costs and tax drag in both windows, with a maximum drawdown no more than 1.5× VTI's.
- Full window (2014-01 to 2025-12): After costs and tax drag it does not beat the benchmark (4.39% against 12.88%), and its maximum drawdown of -37.35% is outside the 1.5× limit (37.35% against a limit of 37.23%).
- 2020–2025 (2020-01 to 2025-12): After costs and tax drag it does not beat the benchmark (7.23% against 14.37%), and its maximum drawdown of -37.35% is outside the 1.5× limit (37.35% against a limit of 37.23%).
This is a fail, and it is a finished result. It is not softened into a partial pass, and it does not earn a forward paper position. Per the charter, the Financier moves to the next hypothesis in the queue (spin-offs; post-earnings-announcement drift).
Full window, 2014-01 to 2025-12
| Measure | Strategy | VTI | Label |
|---|---|---|---|
| Annualised return, before costs | 8.38% | 12.88% | sourced |
| Annualised return, after 0.2%/side costs | 6.75% | 12.88% | assumed cost |
| Annualised return, after costs and tax drag | 4.39% | 12.88% | assumed 35% short-term |
| Annualised volatility | 21.06% | 14.93% | sourced |
| Maximum drawdown | -37.35% | -24.82% | sourced |
| Drawdown peak → trough | 2021-02 → 2023-10 | 2021-12 → 2022-09 | sourced |
| Hit rate of individual positions | 46.31% | — | sourced |
| Completed positions | 4906 | — | sourced |
| Average positions held | 116 | 1 | sourced |
| Annual turnover, one-way | 380.91% | 0% | sourced |
2020 to 2025 alone, so a recent-regime failure is visible
| Measure | Strategy | VTI | Label |
|---|---|---|---|
| Annualised return, before costs | 12.80% | 14.37% | sourced |
| Annualised return, after 0.2%/side costs | 11.12% | 14.37% | assumed cost |
| Annualised return, after costs and tax drag | 7.23% | 14.37% | assumed 35% short-term |
| Annualised volatility | 25.64% | 17.67% | sourced |
| Maximum drawdown | -37.35% | -24.82% | sourced |
| Drawdown peak → trough | 2021-02 → 2023-10 | 2021-12 → 2022-09 | sourced |
| Hit rate of individual positions | 45.36% | — | sourced |
| Completed positions | 2740 | — | sourced |
| Average positions held | 129 | 1 | sourced |
| Annual turnover, one-way | 375.67% | 0% | sourced |
How this was built
- Fundamentals: the SEC Financial Statement Data Sets,
https://www.sec.gov/dera/data/financial-statement-data-sets(sourced). Operating income is the as-reported trailing-twelve-month figure (qtrs=4), taken from the most recent filing whose filing date precedes the ranking date. - Market cap = shares outstanding × month-end close (sourced for both inputs). Shares come from the same filings, on the same filing-date basis.
- The universe averaged 1987 names a month passing the $500M screen, so the top decile was about 199 names, of which 116 survived the quality filter and were held (sourced).
- The quality filter dropped an average of 82.3 names a month for negative operating income. A further 24.2 a month had no filed operating-income figure; the rule drops names that are negative, and a missing figure is not negative, so those were held. That is the rule applied as written rather than tightened after the fact.
- CIK-to-ticker mapping is the SEC's current
company_tickers.json, so companies that have since changed ticker or delisted are mismatched or missing. This is the survivorship bias again, entering through the universe rather than the prices.
H3 — Spin-offs (queue item 3; pre-registered 2026-09-19, sourcing verified 2026-09-26, parent resolution tested 2026-10-03, not yet run)
The rule, fixed before any data is looked at: identify spin-off separations from SEC Form 10-12B / 10-12B-A registration statements (filed under Exchange Act Section 12(b) to list a new entity's common stock after distribution to the parent's shareholders). At completion (the new entity's first full calendar month of regular-way trading), buy the spinoff and its parent equal-weighted at the next month's open; hold 12 months; rebalance monthly across the current cohort — the same mechanic H1 already uses, reused rather than reinvented.
Basis, verified 2026-09-26 against the primary PDFs (read directly with pypdf after WebFetch's summarizer again failed to extract the text layer): Cusatis, Miles & Woolridge, "Restructuring through spinoffs: The stock market evidence," Journal of Financial Economics 33(3), 1993, pp. 293–311. Corrects an error carried from last run's secondary-source summary: the sample is 146 spinoffs from 141 parents, not 161. Spinoffs' mean matched-firm-adjusted return: -1.0% (6mo, n.s.), 4.5% (12mo, n.s.), 25.0% (24mo, sig. 5%), 33.6% (36mo, sig. 5%). Parents (n=131): 6.8%, 12.5%, 26.7%, 18.1% over the same horizons. The abnormal return is concentrated in the 21 of 146 later taken over (61.3% 24-month return vs. 18.9% for the other 125, not significant) — a rule that can't pick the takeover targets in advance is buying a lot of names that contributed roughly nothing.
A second, better-matched paper was found and read in full this run: McConnell, Sibley & Xu, "The Stock Price Performance of Spin-Off Subsidiaries, Their Parents, and the Spin-Off ETF, 2001–2013," Journal of Portfolio Management 42(1), Fall 2015, pp. 143–152 — verified, fetched free from the author's Purdue faculty page. Sample: 146 spin-offs, 2001–2012. Subsidiaries held 22 months: 1.26%/month raw, 0.72%/month excess (cumulative 17.1%) over a size/book-to-market benchmark. Parents held 15 months: 0.85%/month raw, 0.27%/month excess (cumulative 3.7%). The same universe, held 22 months, returned 22.2% annualized vs. 17.4% for the Guggenheim Spin-off ETF (CSD) — the passive version underperformed the paper's own hand-built universe. The "McConnell & Ovtchinnikov" citation from last run is a different, still-unread SSRN paper (abstract_id=569283) — flagged so the two are not conflated; low priority now.
Data source, checked live this run: SEC EDGAR's free full-text search, https://efts.sec.gov/LATEST/search-index, no account, no key. Correction: last run's "437 filings for 2014" counted indexed sub-documents (exhibits), not filings. The new bin/backtest/h3_filings.py dedupes by accession number and finds 188 distinct 10-12B/10-12B-A filings for 2014, collapsing by filer CIK to 50 distinct spinoff registrants — the real population size, in line with published spinoff-frequency estimates. The four names verified by hand last run are all among the 50.
Parent resolution: built and tested on the full 2014 cohort, with disclosed false-positive and false-negative modes. bin/backtest/h3_distribution.py searches for 8-Ks naming each registrant exactly, within 270 days of its last registration amendment, filtered to item 2.01 ("Completion of Acquisition or Disposition of Assets") filed by a different CIK — the item both of last run's worked examples shared. Result: 23/50 resolved to one parent, 4/50 ambiguous, 23/50 unresolved. The 23 include every name already verified by hand (CareTrust REIT→Ensign Group, Theravance Biopharma→Theravance Inc, Kimball Electronics→Kimball International, Communications Sales & Leasing→Windstream Holdings) plus others the method found unaided (ONE Gas→ONEOK, Lands' End→Sears Holdings, Keysight→Agilent, Halyard Health→Kimberly-Clark, NOW Inc→National Oilwell Varco, KLX→B/E Aerospace, Vectrus→Exelis).
Spot-checking against known history found real errors in both directions — "resolved" is not "verified." A confirmed false positive: the quoted search for "Time Inc." matched an unrelated company (Wireless Ronin Technologies) whose 8-K merely mentions the name in passing text with an incidental item-2.01 tag; the real Time Inc. spinoff was from Time Warner. A confirmed false negative, explaining part of the "unresolved" bucket: Knowles Corp (really spun off from Dover Corporation, March 2014) came back "no" only because the registrant's short display name doesn't exact-match Dover's 8-K text, which uses the full legal name "Knowles Corporation" — re-querying with the full name finds Dover's 8-K immediately. Neither fix is built yet. h3_signal.py and run_h3.py remain deliberately unwritten: feeding a signal file from a method with a known, uncorrected false-positive mode would pollute the eventual backtest. A paid dataset (sec-api.io) would shortcut parts of this but requires an account and is out of reach under RULES.md rule 4. No forward position opens on this until H3 is actually run and passes the decision rule in both windows.
What was actually downloaded, and how the prices were checked
The full 2014–2025 window was obtained; nothing had to be cut back to 2018–2025. 96 SEC bulk files downloaded, none failed — 48 quarters of the Form 3/4/5 datasets and 48 of the Financial Statement Data Sets, plus 12 further quarters of Form 3/4/5 covering 2011–2013, which the routine-buyer test needs in order to look three years back from the start of the window. About 5 GB, cached under data/.
The SEC refuses anonymous clients outright: every request without a declared User-Agent came back 403. It is sent one that says what this is, and it carries no real person's address — nothing in this repo sends Wes's email anywhere.
Monthly price bars are this repo's own aggregation of Yahoo's daily data, so they were cross-checked against yfinance on 74 randomly drawn series across four samples. Month-over-month returns agreed to within 0.5% on almost every series and within 1.5% on the worst; one series in one sample disagreed by 30% in a single month, which is the signature of a corporate action the two endpoints adjust differently rather than of an aggregation bug. This check cannot catch an error in Yahoo itself — both sides come from the same place.
One thing worth saying about the windows
Both hypotheses were tested over 2014–2025. Both source papers were published before that: Jegadeesh & Titman in 1993, Cohen, Malloy & Pomorski in 2012. A published anomaly that stops working after it is published is the ordinary case, not a surprising one — once a pattern is in a journal, money moves toward it until it is no longer worth moving. So a fail here is not evidence the papers were wrong about their own windows, and it is not an excuse either: what Wes can trade is the future, and the most recent decade is the closest thing to it that exists.
What this does and does not settle
It settles what these two rules would have produced on this data, with these costs and these biases. It does not settle whether disciplined research beats the index — two mechanical rules are not research, and that was always the point of running them first: they are the cheap, falsifiable version of the question, and they answer in days rather than in a year.
It also does not settle what Wes should do with his money. Nothing here is advice, nothing here was sent to anyone, and no position was taken.
Reproducing this
The code is in bin/backtest/ and the data it caches is in data/, which is git-ignored because it is several gigabytes of re-downloadable public files. In order:
python3 bin/backtest/venv_install.py # deps into /srv/fleet/.venv
python3 bin/backtest/fetch_sec.py --from 2014 --to 2025
python3 bin/backtest/fetch_sec.py --from 2011 --to 2013 --only form345
python3 bin/backtest/h1_signal.py --history-from 2011
python3 bin/backtest/fetch_prices.py data/out/h1_tickers.txt
python3 bin/backtest/h2_fundamentals.py
python3 bin/backtest/h2_universe.py
python3 bin/backtest/fetch_prices.py data/out/h2_tickers.txt
python3 bin/backtest/run_h1.py
python3 bin/backtest/run_h2.py
python3 bin/backtest/report.py