Insider Form 4
Two things decide whether a Form 4 study is real, and the deadline is neither of them.
The deadline is two business days, so decay barely costs you anything. What costs you is
(1) letting grants, option exercises and tax withholdings into the portfolio alongside
actual purchases, and (2) filling at the close of a day whose filing was accepted at
20:30 ET.
Everything marked ✅ Measured is printed by scripts/form4.py (numpy + pandas, seed
20260910, about 5 s, no network). Its filings are synthetic and seeded - the code mix
and the lag distribution are set by the script and measured back, not a tally of EDGAR.
Everything marked ✅ source-verified was read at the URL given, on 2026-09-10.
1. ✅ The rule, at primary sources
| Fact |
Value |
Source, read 2026-09-10 |
| Deadline |
"before the end of the second business day following the day on which the subject transaction has been executed" |
15 U.S.C. 78p(a)(2)(C) (govinfo.gov) and 17 CFR 240.16a-3(g)(1) (eCFR) |
| Deferred execution date |
for a Rule 10b5-1(c) plan trade where "the reporting person does not select the date of execution", and for a 16b-3(f) discretionary transaction, the deemed execution date is the date the broker or administrator "notifies the reporting person" |
17 CFR 240.16a-3(g)(2), (g)(3) |
| 🔑 The cap nobody quotes |
16a-3(g)(4): if notification is later than the third business day after the trade, the execution date is deemed to be that third business day. Outer limit T+5 business days. |
17 CFR 240.16a-3(g)(4) |
| Gifts |
16a-3(g)(1) as amended now covers "dispositions by bona fide gifts" on Form 4. They used to sit on the Form 5, due up to 45 days after fiscal year end. |
17 CFR 240.16a-3(g)(1) |
| Filing window |
a filing after 17:30 ET is "deemed filed as of the next business day", but a Form 3, 4 or 5 "commencing on or before 10 p.m. Eastern" is "deemed filed on the same business day" |
17 CFR 232.13(a)(2) and (a)(4) |
| Same 10 p.m. treatment |
⚠️ (a)(4) is not Forms 3/4/5 only - it also names Schedule 14N, Form 144 and Schedules 13D/13G |
17 CFR 232.13(a)(4) |
| Format |
Ownership XML, submission types "3, 4, 5, 3/A, 4/A, and 5/A"; a live form4.xml on EDGAR returns Content-Type: text/xml, root <ownershipDocument> |
EDGAR Ownership XML Technical Specification; an Archives fetch |
| Access |
"Current max request rate: 10 requests/second", and "Please declare your user agent in request headers" |
sec.gov/os/accessing-edgar-data |
🚨 filingDate is a date; the 22:00 ET window means it is not an availability
timestamp. Use acceptanceDateTime and the availability rule in
../../../fin-core/skills/research-integrity-guards/SKILL.md §2a - which already flags
22:00 ET for Forms 3/4/5 specifically. §5 below prices the difference.
2. ✅ The 20 transaction codes, in the SEC's own groups
Form 4 General Instruction 8 / sec.gov/edgar/searchedgar/ownershipformcodes.html,
read 2026-09-10. There are twenty, not the eighteen or nineteen usually listed - the
one most often dropped is V.
| group |
codes |
the ones that decide a signal |
| General |
P, S, V |
P "Open market or private purchase..."; S "Open market or private sale..."; V "Transaction voluntarily reported earlier than required" |
| Rule 16b-3 |
A, D, F, I, M |
A "Grant, award or other acquisition pursuant to Rule 16b-3(d)"; F "Payment of exercise price or tax liability by delivering or withholding securities..."; M "Exercise or conversion of derivative security exempted pursuant to Rule 16b-3" |
| Derivative ("Except for transactions exempted pursuant to Rule 16b-3") |
C, E, H, O, X |
X "Exercise of in-the-money or at-the-money derivative security"; O the out-of-the-money version; C "Conversion of derivative security" |
| Other 16(b) Exempt and Small Acquisition |
G, L, W, Z |
G "Bona fide gift"; W "...by will or the laws of descent and distribution" |
| Other |
J, K, U |
J "Other acquisition or disposition (describe transaction)"; K is an equity swap and is appended to another code, "e.g., 'S/K' or 'P/K.'" |
🚨 The code is not the direction. I, J, G, W, Z, K and V have no direction in their own
SEC description - I is literally "resulting in acquisition or disposition". Direction is a
separate acquired/disposed flag on each transaction line. A signal keyed on the code string
alone is guessing on seven of twenty codes.
⚠️ "Exempt" is a heading, not a field. Only G, L, W and Z sit under a heading containing
the word; A, D, F, I and M sit under "Rule 16b-3 Transaction Codes", and Rule 16b-3 is the
exemptive rule. C, E, H, O and X are headed "Except for transactions exempted pursuant to
Rule 16b-3".
⚠️ 🔑 V is a selection effect, not a curiosity. A filing marked V arrived before it had
to. Any lag distribution you measure is a mixture of on-time filers, early filers and the
deferred branch, and V is the only one of the three the form labels.
⚠️ Which codes carry information - academic, not SEC
🔴 No SEC publication says any code is informative. Do not attribute it to them. The
finding is academic and it is ⚠️ secondhand here:
- Lakonishok & Lee, "Are Insider Trades Informative?", Review of Financial Studies
14(1), 2001, 79-111: "informativeness of insiders' activities is coming from purchases,
while insider selling appears to have no predictive ability." NYSE/AMEX/Nasdaq, 1975-1995.
- Jeng, Metrick & Zeckhauser, Review of Economics and Statistics 85(2), 2003,
453-471. From the NBER working-paper abstract (w6913): the purchase portfolio "earns
abnormal returns of about 40 basis points per month"; "The sale portfolio does not earn
abnormal returns." ⚠️ The widely quoted ">6% per year" figure from the published version
could not be reached at a primary source; use the 40 bp/month number or verify it.
- Cohen, Malloy & Pomorski, "Decoding Inside Information", Journal of Finance 67(3),
2012, 1009-1043: "opportunistic insider trades yields value-weight abnormal returns of
82 basis points per month, while the abnormal returns associated with routine traders are
essentially zero." 🚨 This is routine-vs-opportunistic timing, not the P-vs-S code
split. Conflating the two is common and wrong.
3. ✅ Measured - the lag and the hour
6,000 seeded filings. draw_lag_and_hour() sets both distributions; these are measured back.
| business days, execution to acceptance |
mean |
median |
p90 |
p95 |
p99 |
max |
|
2.50 |
2 |
4 |
5 |
16 |
49 |
✅ Measured: 75.2% inside the two-business-day deadline, 96.8% inside the T+5 outer
limit that 16a-3(g)(4) implies for the deferred branch, 3.2% past it. And the hour:
✅ Measured: 62.1% of filings accepted at or after the 16:00 close, 50.7% after 17:30,
latest 22.0 ET.
⚠️ Those two percentages are the generator's, not EDGAR's - measure your own feed's. What
is not a choice is the shape: the deadline is two business days rather than 45, and the
ownership-form window runs six hours past the closing bell.
4. ✅ Measured - the A/B, and where the gap actually comes from
Long the purchased name for 63 trading days, short the market, entered at the close of the
key day. Open-market purchases only (code P). Seed 20260910:
| key |
Sharpe |
ann. return |
ann. vol |
gap |
| transaction-date (look-ahead) |
1.869 |
13.58% |
7.26% |
+0.768 |
| filing day's close (look-ahead) |
1.631 |
11.92% |
7.31% |
+0.529 |
| first tradeable session (honest) |
1.102 |
7.98% |
7.25% |
- |
Now decompose it. announce_share is how much of each purchase's alpha arrives as a jump on
the first session the market can trade the filing, rather than as drift from the execution
date. 10 seeds per row:
| announce share |
transaction key |
first tradeable |
gap |
retained |
| 0% |
1.628 |
1.436 |
0.191 |
88.2% |
| 20% |
1.624 |
1.138 |
0.486 |
70.1% |
| 40% |
1.616 |
0.840 |
0.776 |
52.0% |
| 60% |
1.605 |
0.542 |
1.064 |
33.7% |
🔑 🚨 At a two-business-day deadline the lag itself is nearly free - 88.2% retained with
no announcement effect at all. The gap in real Form 4 data is not the deadline; it is the
market's reaction to the filing, which happens on the session you enter at the close of.
This is the opposite diagnosis from ../congressional-trading-disclosures/SKILL.md, where a
45-day lag destroys the signal by decay. Same look-ahead, different mechanism, and the fix
is different: for Form 4 you cannot buy the pop back by trading faster on a daily bar.
⚠️ The announcement share is a parameter, not a measurement of the market. Estimate your own
by comparing the filing-day return of code-P names against a matched sample.
5. 🚨 Measured - what filling at the filing day's close is worth
| fill |
Sharpe |
sd |
| filing day's close, whatever hour it was accepted |
1.272 |
0.351 |
| first tradeable session |
0.840 |
0.363 |
✅ Measured: 0.432 of Sharpe, bought entirely with the 61.3% of filings accepted after
the bell. This is the 90-minute post-close leak that
../../../fin-core/skills/fundamental-and-macro-data/SKILL.md §2.1 describes for periodic
reports, except the ownership-form window makes it a six-hour one.
6. 🚨 Measured - the code filter is the whole game
All four signals honestly keyed to the first tradeable session, same filings, same names,
same 63-day hold. 10 seeds. Only the code filter differs:
| signal |
Sharpe |
sd |
% of filings used |
% of those informed |
| every Form 4 acquisition (P, A, M, C, X) |
0.264 |
0.349 |
53.1% |
23.0% |
| "net insider buying" (acquisitions minus dispositions) |
0.196 |
0.284 |
96.0% |
12.5% |
| purchases and sales (P long, S short) |
0.453 |
0.339 |
36.2% |
32.5% |
| open-market purchases only (P) |
0.840 |
0.363 |
12.1% |
100% |
✅ Measured spread: 0.644 of Sharpe from the filter alone. 🔑 The most inclusive signal
is the worst one. "Net insider buying" reads 96% of filings and only 12.5% of what it reads
is a decision - the rest is the issuer's calendar: annual grants (A), the exercises that
follow them (M) and the shares withheld to pay the tax on them (F). Shorting F is shorting a
payroll deduction.
⚠️ In this generator only code P carries alpha, which is the strongest possible version of
§2's academic finding. If sales carried information the P-and-S row would rise. The ordering
that does not depend on that choice is inclusive < selective, because the denominator
is real: grants and exercises outnumber purchases and they are scheduled, so they also
cluster in calendar time, which concentrates the naive portfolio.
7. Traps
- 🚨 Keying on the execution date. §4, +0.768 of Sharpe on one seed.
- 🚨 Filling at the filing day's close. §5, +0.432. Use
acceptanceDateTime, convert to
exchange local time, and if it is at or after 16:00 the earliest bar is the next session.
- 🚨 No code filter, or a filter built from the code string alone. §6, and §2: seven of
twenty codes carry no direction in their own description.
- 🚨 Treating F as a sale. It is tax withholding on vesting. So is D (disposition back to
the issuer). Neither is a view.
- 🚨 Treating M as a purchase. An option exercise is a conversion, usually on a schedule,
and is very often paired with an immediate S on the same form.
- 🚨 Counting shares instead of reading the A/D flag. A single Form 4 routinely carries an
M, an F and an S on three lines with opposite directions. Summing "shares transacted" across
them produces a number that means nothing.
- ⚠️ 10b5-1 plan trades are pre-committed. §1's deferred branch exists because the insider
did not choose the date. A plan trade is close to the "routine" category in Cohen/Malloy/
Pomorski; treating it as a view is what that paper says not to do.
- 🚨 Amendments (4/A) backfilled onto the original date. A 4/A is a new publication with
its own acceptance timestamp. Same rule as a restated 10-K.
- 🚨 Survivorship in the issuer universe. Insider buying concentrates in small, distressed
names - exactly the ones a current-constituents universe has dropped.
../../../fin-core/skills/research-integrity-guards/SKILL.md §1.
- ⚠️ Business days are not trading days. The deadline counts business days; the exchange
calendar has holidays the federal one does not, and vice versa. This script treats them as
the same and says so.
8. Scripts and where this sits
scripts/form4.py - the transcribed deadline, deferred branch, filing window and all 20
codes in TRANSACTION_CODES; draw_lag_and_hour() and lag_summary(); make_panel() with
announce_share and scheduled-code clustering; portfolio(), run_ab(), announce_split(),
post_close_cost(), signal_sides() and code_filter_table(). numpy + pandas, seed
20260910, about 5 s, no network, no file writes, no scraper.
../congressional-trading-disclosures/SKILL.md - the same A/B with a 45-day deadline,
where the gap comes from decay rather than the announcement, and where the amounts are
disclosed as brackets.
../institutional-13f/SKILL.md - the same A/B where the clock starts at a quarter end.
../social-and-influencer-feeds/SKILL.md - the same trap with no filing at all.
../../../fin-core/skills/fundamental-and-macro-data/SKILL.md §1 and §2.1 - the EDGAR
endpoints, the required User-Agent, the 10 req/s limit, and the period/filed/accepted
distinction. edgartools is the library it recommends for typed Form 4 objects.
../../../fin-core/skills/research-integrity-guards/SKILL.md §2a - the availability rule,
which already names the 22:00 ET ownership-form cutoff.
../../../fin-libraries/skills/lib-edgartools/SKILL.md - the per-library page.
../../../fin-core/skills/backtest-validation/SKILL.md - four code filters is four trials,
and a Sharpe of 0.840 with sd 0.363 across seeds needs an interval, not a decimal.
1---2name: insider-form-43description: >-4 Filter Form 4 to open-market purchases by transaction code, then key the signal to the5 first session that can trade the acceptance timestamp. TRIGGER - Form 4, Forms 3/4/5,6 Section 16, insider buying, insider selling, corporate insider trades, "officer bought7 shares", transaction codes P S A M F G X C, code P vs code A, 10b5-1 plan, EDGAR ownership8 XML, ownershipDocument, transactionCode, two business days, "insider trading signal",9 "should I follow insider buys", openinsider-style screens, "why does my insider backtest10 look amazing", stripping option exercises and tax withholding out of insider data. SKIP11 for congressional trades and the STOCK Act (congressional-trading-disclosures), for fund12 holdings and 13F staleness (institutional-13f), for social posts and influencer panels13 (social-and-influencer-feeds), for EDGAR endpoints, edgartools and XBRL fundamentals14 (fundamental-and-macro-data), and for the general availability rule15 (research-integrity-guards).16license: MIT17metadata:18 source: https://skillmd.com/skills/howard-lynn-ye/insider-form-419 version: "0.1.0"20 verified_on: "2026-09-10"21---2223# Insider Form 42425**Two things decide whether a Form 4 study is real, and the deadline is neither of them.**26The deadline is two business days, so decay barely costs you anything. What costs you is27(1) letting grants, option exercises and tax withholdings into the portfolio alongside28actual purchases, and (2) filling at the close of a day whose filing was accepted at2920:30 ET.3031Everything marked ✅ Measured is printed by `scripts/form4.py` (numpy + pandas, seed3220260910, about 5 s, no network). Its filings are **synthetic and seeded** - the code mix33and the lag distribution are set by the script and measured back, not a tally of EDGAR.34Everything marked ✅ source-verified was read at the URL given, on 2026-09-10.3536## 1. ✅ The rule, at primary sources3738| Fact | Value | Source, read 2026-09-10 |39|---|---|---|40| **Deadline** | *"before the end of the second business day following the day on which the subject transaction has been executed"* | **15 U.S.C. 78p(a)(2)(C)** (`govinfo.gov`) and **17 CFR 240.16a-3(g)(1)** (eCFR) |41| Deferred execution date | for a Rule 10b5-1(c) plan trade where *"the reporting person does not select the date of execution"*, and for a 16b-3(f) discretionary transaction, the deemed execution date is the date the broker or administrator *"notifies the reporting person"* | 17 CFR 240.16a-3(g)(2), (g)(3) |42| 🔑 **The cap nobody quotes** | 16a-3(g)(**4**): if notification is later than the **third** business day after the trade, the execution date is deemed to be that third business day. **Outer limit T+5 business days.** | 17 CFR 240.16a-3(g)(4) |43| Gifts | 16a-3(g)(1) as amended now covers *"dispositions by bona fide gifts"* on Form 4. They used to sit on the Form 5, due up to 45 days after fiscal year end. | 17 CFR 240.16a-3(g)(1) |44| **Filing window** | a filing after **17:30 ET** is *"deemed filed as of the next business day"*, **but** a Form 3, 4 or 5 *"commencing on or before 10 p.m. Eastern"* is *"deemed filed on the same business day"* | **17 CFR 232.13(a)(2)** and **(a)(4)** |45| Same 10 p.m. treatment | ⚠️ (a)(4) is **not** Forms 3/4/5 only - it also names **Schedule 14N, Form 144 and Schedules 13D/13G** | 17 CFR 232.13(a)(4) |46| Format | Ownership XML, submission types *"3, 4, 5, 3/A, 4/A, and 5/A"*; a live `form4.xml` on EDGAR returns `Content-Type: text/xml`, root `<ownershipDocument>` | EDGAR Ownership XML Technical Specification; an Archives fetch |47| Access | *"Current max request rate: 10 requests/second"*, and *"Please declare your user agent in request headers"* | `sec.gov/os/accessing-edgar-data` |4849🚨 **`filingDate` is a date; the 22:00 ET window means it is not an availability50timestamp.** Use `acceptanceDateTime` and the availability rule in51`../../../fin-core/skills/research-integrity-guards/SKILL.md` §2a - which already flags5222:00 ET for Forms 3/4/5 specifically. §5 below prices the difference.5354## 2. ✅ The 20 transaction codes, in the SEC's own groups5556Form 4 General Instruction 8 / `sec.gov/edgar/searchedgar/ownershipformcodes.html`,57read 2026-09-10. There are **twenty**, not the eighteen or nineteen usually listed - the58one most often dropped is **V**.5960| group | codes | the ones that decide a signal |61|---|---|---|62| **General** | P, S, **V** | **P** *"Open market or private purchase..."*; **S** *"Open market or private sale..."*; **V** *"Transaction voluntarily reported earlier than required"* |63| **Rule 16b-3** | A, D, F, I, M | **A** *"Grant, award or other acquisition pursuant to Rule 16b-3(d)"*; **F** *"Payment of exercise price or tax liability by delivering or withholding securities..."*; **M** *"Exercise or conversion of derivative security exempted pursuant to Rule 16b-3"* |64| **Derivative** (*"Except for transactions exempted pursuant to Rule 16b-3"*) | C, E, H, O, X | **X** *"Exercise of in-the-money or at-the-money derivative security"*; **O** the out-of-the-money version; **C** *"Conversion of derivative security"* |65| **Other 16(b) Exempt and Small Acquisition** | G, L, W, Z | **G** *"Bona fide gift"*; W *"...by will or the laws of descent and distribution"* |66| **Other** | J, K, U | J *"Other acquisition or disposition (describe transaction)"*; K is an equity swap and is **appended** to another code, *"e.g., 'S/K' or 'P/K.'"* |6768🚨 **The code is not the direction.** I, J, G, W, Z, K and V have no direction in their own69SEC description - I is literally *"resulting in acquisition or disposition"*. Direction is a70separate acquired/disposed flag on each transaction line. A signal keyed on the code string71alone is guessing on seven of twenty codes.7273⚠️ **"Exempt" is a heading, not a field.** Only G, L, W and Z sit under a heading containing74the word; A, D, F, I and M sit under *"Rule 16b-3 Transaction Codes"*, and Rule 16b-3 is the75exemptive rule. C, E, H, O and X are headed *"Except for transactions exempted pursuant to76Rule 16b-3"*.7778⚠️ 🔑 **V is a selection effect, not a curiosity.** A filing marked V arrived *before* it had79to. Any lag distribution you measure is a mixture of on-time filers, early filers and the80deferred branch, and V is the only one of the three the form labels.8182### ⚠️ Which codes carry information - academic, not SEC8384🔴 **No SEC publication says any code is informative.** Do not attribute it to them. The85finding is academic and it is ⚠️ secondhand here:8687- **Lakonishok & Lee**, *"Are Insider Trades Informative?"*, **Review of Financial Studies88 14(1), 2001, 79-111**: *"informativeness of insiders' activities is coming from purchases,89 while insider selling appears to have no predictive ability."* NYSE/AMEX/Nasdaq, 1975-1995.90- **Jeng, Metrick & Zeckhauser**, *Review of Economics and Statistics* **85(2), 2003**,91 453-471. From the NBER working-paper abstract (w6913): the purchase portfolio *"earns92 abnormal returns of about 40 basis points per month"*; *"The sale portfolio does not earn93 abnormal returns."* ⚠️ The widely quoted ">6% per year" figure from the published version94 could **not** be reached at a primary source; use the 40 bp/month number or verify it.95- **Cohen, Malloy & Pomorski**, *"Decoding Inside Information"*, **Journal of Finance 67(3),96 2012**, 1009-1043: *"opportunistic insider trades yields value-weight abnormal returns of97 82 basis points per month, while the abnormal returns associated with routine traders are98 essentially zero."* 🚨 **This is routine-vs-opportunistic *timing*, not the P-vs-S code99 split.** Conflating the two is common and wrong.100101## 3. ✅ Measured - the lag and the hour1021036,000 seeded filings. `draw_lag_and_hour()` sets both distributions; these are measured back.104105| business days, execution to acceptance | mean | median | p90 | p95 | p99 | max |106|---|---|---|---|---|---|---|107| | **2.50** | **2** | 4 | 5 | 16 | 49 |108109✅ Measured: **75.2%** inside the two-business-day deadline, **96.8%** inside the T+5 outer110limit that 16a-3(g)(4) implies for the deferred branch, **3.2%** past it. And the hour:111112✅ Measured: **62.1%** of filings accepted at or after the 16:00 close, **50.7%** after 17:30,113latest 22.0 ET.114115⚠️ Those two percentages are the generator's, not EDGAR's - **measure your own feed's**. What116is *not* a choice is the shape: the deadline is two business days rather than 45, and the117ownership-form window runs six hours past the closing bell.118119## 4. ✅ Measured - the A/B, and where the gap actually comes from120121Long the purchased name for 63 trading days, short the market, entered at the **close of the122key day**. Open-market purchases only (code P). Seed 20260910:123124| key | Sharpe | ann. return | ann. vol | gap |125|---|---|---|---|---|126| **transaction-date** (look-ahead) | **1.869** | 13.58% | 7.26% | **+0.768** |127| **filing day's close** (look-ahead) | 1.631 | 11.92% | 7.31% | +0.529 |128| **first tradeable session** (honest) | **1.102** | 7.98% | 7.25% | - |129130Now decompose it. `announce_share` is how much of each purchase's alpha arrives as a jump on131the first session the market can trade the filing, rather than as drift from the execution132date. 10 seeds per row:133134| announce share | transaction key | first tradeable | gap | retained |135|---|---|---|---|---|136| **0%** | 1.628 | **1.436** | **0.191** | **88.2%** |137| 20% | 1.624 | 1.138 | 0.486 | 70.1% |138| **40%** | 1.616 | **0.840** | **0.776** | **52.0%** |139| 60% | 1.605 | 0.542 | 1.064 | 33.7% |140141🔑 🚨 **At a two-business-day deadline the lag itself is nearly free - 88.2% retained with142no announcement effect at all.** The gap in real Form 4 data is not the deadline; it is the143market's reaction to the filing, which happens on the session you *enter at the close of*.144This is the opposite diagnosis from `../congressional-trading-disclosures/SKILL.md`, where a14545-day lag destroys the signal by decay. **Same look-ahead, different mechanism, and the fix146is different: for Form 4 you cannot buy the pop back by trading faster on a daily bar.**147148⚠️ The announcement share is a parameter, not a measurement of the market. Estimate your own149by comparing the filing-day return of code-P names against a matched sample.150151## 5. 🚨 Measured - what filling at the filing day's close is worth152153| fill | Sharpe | sd |154|---|---|---|155| **filing day's close**, whatever hour it was accepted | **1.272** | 0.351 |156| **first tradeable session** | **0.840** | 0.363 |157158✅ Measured: **0.432 of Sharpe**, bought entirely with the **61.3%** of filings accepted after159the bell. This is the 90-minute post-close leak that160`../../../fin-core/skills/fundamental-and-macro-data/SKILL.md` §2.1 describes for periodic161reports, except the ownership-form window makes it a **six-hour** one.162163## 6. 🚨 Measured - the code filter is the whole game164165All four signals honestly keyed to the first tradeable session, same filings, same names,166same 63-day hold. 10 seeds. Only the code filter differs:167168| signal | Sharpe | sd | % of filings used | % of those informed |169|---|---|---|---|---|170| every Form 4 acquisition (P, A, M, C, X) | **0.264** | 0.349 | 53.1% | 23.0% |171| "net insider buying" (acquisitions minus dispositions) | **0.196** | 0.284 | 96.0% | 12.5% |172| purchases and sales (P long, S short) | 0.453 | 0.339 | 36.2% | 32.5% |173| **open-market purchases only (P)** | **0.840** | 0.363 | **12.1%** | **100%** |174175✅ Measured spread: **0.644 of Sharpe from the filter alone.** 🔑 **The most inclusive signal176is the worst one.** "Net insider buying" reads 96% of filings and only 12.5% of what it reads177is a decision - the rest is the issuer's calendar: annual grants (A), the exercises that178follow them (M) and the shares withheld to pay the tax on them (F). Shorting F is shorting a179payroll deduction.180181⚠️ In this generator only code P carries alpha, which is the strongest possible version of182§2's academic finding. If sales carried information the P-and-S row would rise. The ordering183that does **not** depend on that choice is *inclusive < selective*, because the denominator184is real: grants and exercises outnumber purchases and they are scheduled, so they also185cluster in calendar time, which concentrates the naive portfolio.186187## 7. Traps188189- 🚨 **Keying on the execution date.** §4, +0.768 of Sharpe on one seed.190- 🚨 **Filling at the filing day's close.** §5, +0.432. Use `acceptanceDateTime`, convert to191 exchange local time, and if it is at or after 16:00 the earliest bar is the next session.192- 🚨 **No code filter, or a filter built from the code string alone.** §6, and §2: seven of193 twenty codes carry no direction in their own description.194- 🚨 **Treating F as a sale.** It is tax withholding on vesting. So is D (disposition back to195 the issuer). Neither is a view.196- 🚨 **Treating M as a purchase.** An option exercise is a conversion, usually on a schedule,197 and is very often paired with an immediate S on the same form.198- 🚨 **Counting shares instead of reading the A/D flag.** A single Form 4 routinely carries an199 M, an F and an S on three lines with opposite directions. Summing "shares transacted" across200 them produces a number that means nothing.201- ⚠️ **10b5-1 plan trades are pre-committed.** §1's deferred branch exists because the insider202 did not choose the date. A plan trade is close to the "routine" category in Cohen/Malloy/203 Pomorski; treating it as a view is what that paper says not to do.204- 🚨 **Amendments (4/A) backfilled onto the original date.** A 4/A is a new publication with205 its own acceptance timestamp. Same rule as a restated 10-K.206- 🚨 **Survivorship in the issuer universe.** Insider buying concentrates in small, distressed207 names - exactly the ones a current-constituents universe has dropped.208 `../../../fin-core/skills/research-integrity-guards/SKILL.md` §1.209- ⚠️ **Business days are not trading days.** The deadline counts business days; the exchange210 calendar has holidays the federal one does not, and vice versa. This script treats them as211 the same and says so.212213## 8. Scripts and where this sits214215`scripts/form4.py` - the transcribed deadline, deferred branch, filing window and all 20216codes in `TRANSACTION_CODES`; `draw_lag_and_hour()` and `lag_summary()`; `make_panel()` with217`announce_share` and scheduled-code clustering; `portfolio()`, `run_ab()`, `announce_split()`,218`post_close_cost()`, `signal_sides()` and `code_filter_table()`. numpy + pandas, seed21920260910, about 5 s, no network, no file writes, no scraper.220221- `../congressional-trading-disclosures/SKILL.md` - the same A/B with a **45-day** deadline,222 where the gap comes from decay rather than the announcement, and where the amounts are223 disclosed as brackets.224- `../institutional-13f/SKILL.md` - the same A/B where the clock starts at a *quarter end*.225- `../social-and-influencer-feeds/SKILL.md` - the same trap with no filing at all.226- `../../../fin-core/skills/fundamental-and-macro-data/SKILL.md` §1 and §2.1 - the EDGAR227 endpoints, the required User-Agent, the 10 req/s limit, and the period/filed/accepted228 distinction. `edgartools` is the library it recommends for typed Form 4 objects.229- `../../../fin-core/skills/research-integrity-guards/SKILL.md` §2a - the availability rule,230 which already names the 22:00 ET ownership-form cutoff.231- `../../../fin-libraries/skills/lib-edgartools/SKILL.md` - the per-library page.232- `../../../fin-core/skills/backtest-validation/SKILL.md` - four code filters is four trials,233 and a Sharpe of 0.840 with sd 0.363 across seeds needs an interval, not a decimal.