INSIDER BACKTEST — PRE-REGISTRATION
(English translation of the internal pre-registration file. Rules identical;
original locked by editorial approval on 2026-08-09, content-freeze commit
8f1543c, BEFORE any results were computed. Post-hoc additions are labeled
as such in the article.)

QUESTION
"If I had copied every open-market insider purchase on the day it was filed,
would I have beaten the market?"

1. DATA
- Source: SEC DERA Insider Transactions Data Sets, 2006Q1-2026Q2 (82 quarterly ZIPs)
- Prices: yfinance daily OHLC, split/dividend adjusted (auto_adjust) + SPY
- Signal: NONDERIV_TRANS, transaction code P, acquisition flag A
- Excluded: option exercises (M), grants (A), derivative transactions, sales
- Schema verified against the first ZIP's own documentation before coding

2. DEDUPLICATION (mandatory - all result counts AFTER this)
- 4/A amendment chains: for the same (issuer CIK, insider CIK, transaction date),
  the LATEST filing supersedes (amendment replaces original, accession-level)
- Same-lot duplicates: exact copies on (issuer CIK, insider CIK, transaction
  date, code, shares, price) collapse to one (latest filing kept)
- Report: row counts before/after - any headline number is the AFTER count

3. MAPPING
- CIK -> ticker -> yfinance; MATCH RATE reported next to every results table
- Ticker priority: DERA ISSUERTRADINGSYMBOL -> SEC company_tickers.json ->
  unmatched tail reported with reason distribution
- Known limit: historical ticker changes cannot be fully resolved; stated openly

4. EXPERIMENT DESIGN
- Entry: OPEN of the first trading day AFTER filed_date. Rationale: most Form 4s
  are filed outside market hours; the copier's first realistic order is
  "buy at next open". Events with no open within 7 days of filing are dropped
  and counted.
- Horizons: 1 / 3 / 6 / 12 calendar months; exit = CLOSE of first trading day
  at or after the horizon end
- Benchmark: SPY (auto_adjust) over the same entry/exit days; reported metric =
  excess return (stock minus SPY)
- Metrics: mean, median, win rate, distribution, event counts
- No p-values (overlapping events are dependent); distributions and medians only

5. SLICES (defined before seeing results - no slices added afterwards)
- All events (baseline)
- Cluster: 3+ DISTINCT insiders with P-buys at the same issuer within 14
  calendar days
- Role: CEO/CFO / other officer / director / 10% owner (multi-role priority:
  CEO/CFO > officer > director > 10%)
- Transaction size (shares x price): <$10K / $10K-100K / $100K-1M / >=$1M
- Liquidity buckets (20 trading days of avg dollar volume before entry):
  <$1M/day / $1M-10M/day / >=$10M/day. NOTE: the template's "market-cap
  buckets" were deliberately REPLACED with liquidity before locking - reliable
  historical market cap is not available from free sources; liquidity is both
  computable and the direct measure of copyability. (Change made BEFORE
  results, pre-lock.)

6. PORTFOLIO SIMULATION ("if I had started with $10,000")
- Holding period: 6 months (middle horizon - chosen in advance, not by results)
- Position: fixed $500; max 20 concurrent; skip signal when out of cash
  (skips counted); one position per ticker at a time; order: filing time;
  dividends included (adjusted close)
- Costs: main result frictionless; sensitivity variant with 0.5% round-trip
  total cost reported (as a nod to micro-cap spreads)

7. KNOWN LIMITS (to be stated openly in the article and any video)
- Survivorship: direction UNKNOWN - missing delisted names hide bankruptcies
  (flattering) AND buyout premiums (unflattering); no one-directional claim
- Mapping tail (see 3) + events dropped by the 7-day entry rule
- Cost model is simple (see 6 sensitivity); real slippage not modeled
- Multiple hypotheses: slices pre-declared; distribution reporting, no p-values

8. COMMITMENT
These rules were locked before results were computed. Content-freeze commit:
8f1543c - editorial lock: 2026-08-09. Any later rule change requires a new
commit with written justification. Post-hoc analyses performed after results
(and labeled as such in the article): sub-$0.01 ghost-price sensitivity;
lot-to-purchase-day robustness collapse; proportional position-sizing variant
of the simulation (equity/20) after a cash-drag flaw was found in the fixed
sizing comparison; slice tables re-run under the ghost filter.