Insider Transaction Filing Signal Research

When to Use

Invoke this when building an equity factor from SEC Form 4 filings — the Section 16(a) statement of changes in beneficial ownership filed by officers, directors and >10% beneficial owners of a US registered class. The engine turns a stream of Form 4 transaction lines into a point-in-time, role-weighted net insider sentiment score in $[-1, +1]$.

Two things make this harder than summing buys and sells, and the module exists for both:

• The trade date is not the information date. Rule 16a-3(g)(1) gives the filer until "the end of the second business day following the day on which the subject transaction has been executed." For plan trades where the insider does not pick the execution date, Rule 16a-3(g)(2)–(4) deem the broker-notification date to be the execution date, capped at the third business day after the trade — so the lawful gap between execution and public filing reaches roughly five business days. Cohen, Malloy & Pomorski measured a median trade-to-report delay of 3 days across 1986–2007. A backtest keyed on transaction_date trades on information that did not exist. Every score here is taken as of a filing_datetime cut-off.

• "Routine" is a property of the trader, not of the checkbox. The 82 bps/month result rests on Cohen, Malloy & Pomorski's timing test — an insider who "placed a trade in the same calendar month for at least three consecutive years" is routine. Their 1986–2007 sample uses no 10b5-1 information at all, and they report that "in unreported results we find very similar differential performance of opportunistic versus routine trades before 2000, suggesting that our results are not driven by trades in these plans." classify_trader_regularity() implements that test.

When NOT to Use

• Outside US Section 16. Form 4 covers officers, directors and >10% beneficial owners of a class registered under Section 12. It has no bearing on UK PDMR notifications (UK MAR Art. 19), EU MAR managers' transactions, SEBI PIT Regulation 7(2) disclosures, or any non-US regime; those have different deadlines, different filers and different fields.

• As a long-history backtest of the 10b5-1 filter. The checkbox became mandatory only for reports filed on or after 1 April 2023 (SEC Release 33-11138). Before that, "the disclosure of a purchase or sale under a Rule 10b5-1 trading arrangement in Forms 4 and 5 is voluntary, resulting in a lack of consistent and comprehensive information about such trades." An unchecked box in 2019 means not disclosed, not not a plan trade. Supply those records as PlanStatus.UNKNOWN and read unknown_plan_status_count before trusting any pre-2023 result.

• As a claim that 10b5-1 trades are uninformative. They are not — see Common Pitfalls. The default exclude_plan_trades=True buys a cleaner opportunistic subsample at the cost of discarding documented signal. That is a research choice, not a free lunch.

• On derivative-only or compensation activity. Grants (A), exercises (M, X), tax withholding (F) and gifts (G) are not directional open-market trades. They are counted and reported, never scored.

• As an insider-trading compliance or surveillance control. This reads public filings for alpha. For MNPI handling and alt-data governance see insider-trading-controls-for-alternative-data-usage; for market-abuse self-detection see wash-trade-and-spoofing-self-detection.

Prerequisites

• Form 4 Table I transaction lines carrying, per line: transaction_code (General Instruction 8), shares, price, transaction_date, and the EDGAR filing_datetime at which the filing was disseminated. The filing timestamp is filing metadata, not a field inside the document — source it from the EDGAR index, not from periodOfReport.
• Reporting-owner relationship as the four independent EDGAR booleans isDirector, isOfficer, isTenPercentOwner, isOther, plus the free-text officerTitle (max 30 characters). These are not mutually exclusive: a founder-CEO on the board sets three of them.
• A determination of whether each P/S line actually executed on-market. Codes P and S read "Open market or private purchase/sale", so the code alone does not establish it.
• A role weight schedule. The bundled default (CEO/CFO 1.0, other officer 0.8, director 0.6, 10% owner 0.3) is illustrative — no published source establishes these values.

Workflow

1. Ingest with both dates, and reject naive timestamps. filing_datetime must be timezone-aware; the engine raises otherwise. EDGAR's Forms 3/4/5 cut-off is 22:00 ET, and a naive instant silently adopts the host clock — which is how look-ahead bias enters a pipeline that looks correct.

2. Evaluate as of an explicit instant. analyze_form4_filings(..., as_of=...) scores only filings with filing_datetime <= as_of. Pass the whole history and roll as_of forward; the filings not yet public are counted in not_yet_public_excluded_count, not silently dropped.

   • Decision point — do not "fix" the lag with a fixed +2-day offset on the transaction date. Two business days is the deadline, not the observed delay, and Rule 16a-3(g)(2)–(4) legally extends it for plan trades. Late filings exist. Use the actual dissemination timestamp and read max_trade_to_file_lag_days to see what the feed really delivered.

3. Decide the Rule 10b5-1 policy explicitly, per sample period.

   • Post-2023-04-01 data: the checkbox is reliable; PlanStatus.PLAN / NON_PLAN are meaningful.
   • Pre-2023-04-01 data: supply PlanStatus.UNKNOWN. Then choose — treat_unknown_plan_status_as_plan=True drops them (conservative, costs most of the sample), False scores them and exposes the contamination through unknown_plan_status_count.
   • Decision point — a regime boundary runs through 1 April 2023. A backtest spanning it is comparing two different datasets. Split the sample there or report the two halves separately.

4. Classify trader regularity from prior years only. classify_trader_regularity(history, classification_year=Y) reads transactions dated strictly before Y, exactly as CMP designate insiders at the beginning of each calendar year. Three labels, not two: an insider who did not trade in each of the preceding years is UNCLASSIFIED, and CMP leave those out of the portfolio rather than calling them opportunistic.

5. Resolve the Section 16 capacity from the flags, not from a role string. The engine takes the highest-weighted applicable tier across isOfficer / isDirector / isTenPercentOwner, with the officer tier read out of the free-text title.

   • Decision point — an unparseable role must not get a middle weight. unknown_role_weight defaults to 0.0, excluding the trade from the score and logging it, rather than guessing. Every such record is counted in unclassified_role_count.

6. Score, then apply the sample floors before reading a direction. Weighted notional is $w \times \text{shares} \times \text{price}$; the score is $S = (\Sigma_{\text{buy}} - \Sigma_{\text{sell}}) / (\Sigma_{\text{buy}} + \Sigma_{\text{sell}}) \in [-1, +1]$.

   • Decision point — the score is scale-free, so a single $1,000 purchase reads +1.00. Set min_total_notional_usd and min_distinct_insiders; below either, the engine returns INSUFFICIENT_DATA rather than a saturated signal built on one trade.

7. Reconcile the report before using it. The exclusion counters plus the scored counts equal filings_supplied exactly. If they don't, records were lost upstream.

Full procedure: see references/workflows.md. Standards reference: see references/standards.md. Printable pre-flight checklist: see assets/checklist.md.

Common Pitfalls

• Assuming Rule 10b5-1 sales carry no signal. The SEC's own economic analysis in Release 33-11138 documents -2.5% six-month industry-adjusted returns after the first sale under plans whose first trade fell within 30 days of adoption, and finds single-trade plans — 49% of the plans studied — "consistently loss-avoiding regardless of cooling-off period," avoiding declines up to -4%. Jagolinzer (2009) found plan participants' "sales systematically follow positive and precede negative firm performance." Filtering plan trades is a subsample choice; treating them as noise by definition is wrong.

• Reading an unchecked 10b5-1 box on a pre-April-2023 filing as NON_PLAN. Disclosure was voluntary until then. The SEC's own footnote on a study using 2003–2006 Form 4 data warns that "trades classified as 'non-10b5-1' trades in the study may include 10b5-1 plan trades."

• Aligning the signal to transaction_date. The market cannot act on a trade before EDGAR disseminates it. This is the look-ahead bias that makes an insider factor backtest well and trade badly, and it is invisible in the equity curve.

• Adding a flat +2 business days instead of using the filing timestamp. Rule 16a-3(g)(2)–(4) legally deems the broker-notification date to be the execution date for plan trades, capped at the third business day after the trade — so a compliant filing can land ~5 business days out. Late filings exist beyond that.

• Treating a single role string as the filer's capacity. isDirector, isOfficer, isTenPercentOwner and isOther are four independent booleans; there is no structured "CEO" field anywhere on Form 4. The title is 30 characters of free text, so "Chairman, CEO & Pres", "Chief Executive Officer" and "C.E.O." are the same person's capacity spelled three ways.

• Matching transaction codes with code == "S". General Instruction 8 directs filers to report equity-swap-linked trades as "S/K" or "P/K". A bare equality test drops every one of them without a trace.

• Treating code P as proof of an open-market trade. The SEC's own definition is "Open market or private purchase of non-derivative or derivative security."

• Double-counting Form 4/A amendments. An amendment restates transaction lines under a new accession number. A feed carrying the original and the amendment counts the same economic trade twice; the engine warns on a repeated filing_id but cannot resolve amendment chains for you.

• Equal-weighting a 10% beneficial owner and a CEO. An outside fund crossing 10% files the same form as an officer with direct operational visibility, and its trade is usually a portfolio decision about itself, not a view on the issuer.

Verification

• Instantiate InsiderFilingSignalEngine(). Audit a CEO open-market purchase (10,000 shares @ $50.00, plan_status=PlanStatus.NON_PLAN, is_officer=True, officer_title="Chief Executive Officer") whose filing_datetime is after as_of ⟹ verify not_yet_public_excluded_count == 1 and signal_classification == INSUFFICIENT_DATA. Advance as_of past the dissemination instant ⟹ verify STRONG_BULLISH_OPPORTUNISTIC_BUY at $S = +1.00$. This is the point-in-time gate; it is the single most important behaviour here.
• Audit a disclosed 10b5-1 sale ⟹ verify routine_10b5_1_filtered_count == 1 and INSUFFICIENT_DATA; re-run with exclude_plan_trades=False ⟹ verify it scores at $S = -1.00$.
• Audit an insider who traded every March for three consecutive prior years ⟹ verify classify_trader_regularity labels them ROUTINE, and that the engine excludes their trades when the labels are supplied.
• Confirm the report reconciles: exclusion counters + scored counts == filings_supplied.
• Run python -m unittest discover -s skills/insider-transaction-filing-signal-research/scripts.

Related Skills

• insider-trading-controls-for-alternative-data-usage
• earnings-call-transcript-nlp-signal-research
• point-in-time-fundamentals-data-joins
• lookahead-bias-elimination
• survivorship-bias-free-universe-construction