DRAFT-002 — High-conviction disclosure drift

A Draft. No standing, no alpha spent, nothing measured.

It trades the disclosure, not the purchase — and the whole document turns on that distinction, so the failed version is written out first.

Why the obvious version fails

The natural formulation is: use 13F to find when funds bought, map those dates to the chart, learn the pattern that precedes a large institutional buy. It is intuitive, it is the version most people build, and it cannot work. Three independent reasons, each sufficient.

1. 13F carries no timing

A 13F is a quarterly snapshot of holdings, not a transaction record, filed within 45 days of quarter end.

purchase        2 January
quarter ends    31 March     +88 days
filing due      15 May       +133 days

What the filing discloses is that the position on 31 March differed from the position on 31 December. When it changed is not in the document. There is a 90-day ambiguity window inside up to 135 days of latency, and the position may have been bought, sold and rebought within it.

So a signal that fires before a large institutional buy is unobtainable from this source by construction. The buy is already four months old. The only thing left to be early to is other people's reaction to a public document that is machine-parsed within seconds of reaching EDGAR.

Three further gaps, each of which quietly inverts the signal:

• Long positions only. A manager long $50m and short $80m of the same name files as a holder.
• A vanished position is ambiguous — sold, or fell below the filing threshold, or ceased to be a 13(f) security.
• 13F/A amendments restate. Reading today's amended figure is reading a document that did not exist at the time — ADR-0004's back-adjustment trap wearing different clothes.

2. Inferring the buy date from price is circular

Since the purchase date is unobserved, the only way to recover it is to infer it — and the evidence available is price and volume. Which produces:

1. Infer the buy date from price and volume.
2. Search for price and volume features that predict the inferred date.
3. Find them.
4. Conclude the chart predicts institutional buying.

The chart was used twice. The label was manufactured from the features that then predict it. This is label leakage, and it is not a risk to be managed — it is a guarantee. The backtest will look excellent and will be measuring its own construction.

It is the vacuous pass in its most persuasive disguise, because the output is populated with real funds, real companies and real charts.

3. The search space cannot be declared

"Look for features that build a signal" is unbounded — any indicator, any lookback, any threshold, any combination, and news features on top. DESIGN-v4 D13 already names this failure: *a search that search_space_size never counts*.

It is also mechanically fatal rather than merely unwise. M6.4 computes spend as configured_test_alpha x search_space_size. There is no number to declare, so registration is refused before anything is spent — correctly. What that formulation describes is feature discovery, a legitimate activity requiring a validation regime this lab does not implement (nested cross-validation, deflated Sharpe, probability of backtest overfitting), and which ADR-0014's forward window cannot retrofit.

News is excluded from this Draft for the same family of reasons: coverage is endogenous to price, so "news predicts moves" usually runs backwards; and free archives cannot supply a reliable Known-at instant, which the glossary requires.

The mechanism, stated honestly

Entry is the filing timestamp — EDGAR's acceptance datetime, a genuine known-at instant. No timing is inferred, so the circularity is gone.

The claim is slow information diffusion under limited attention. Thousands of filings land in a two-day cluster at the deadline. Attention is finite. Positions held by less-followed managers, in less-covered names, diffuse into price over days to weeks rather than instantly, and a high-conviction position — one that is a large share of the manager's own portfolio — carries more information than an incremental add.

This mechanism is weaker than DRAFT-001's, and the difference should not be glossed. DRAFT-001 rests on compulsion: an index fund must sell. This rests on inattention, which is a real and documented phenomenon (post-earnings-announcement drift is the canonical case) but has nobody forced to be on the other side. Compulsion is the stronger prior. If only one of these can be afforded, that asymmetry should decide it.

Prior evidence cuts against. Holdings-cloning ETFs were launched commercially in the 2010s and broadly underperformed or closed. That is not proof the effect is absent, but it is strong evidence the obvious form is arbitraged, and it should be weighed before alpha is spent, not after.

Both interpretations

If real. Some managers hold persistent skill, and disclosure of their high-conviction positions is processed slowly enough for a retail participant to act. The effect should be larger in less-covered names and among less-followed managers — a directional prediction, declared here, tested as a pre-declared subgroup.

If not, four distinct worlds, and they are not equally interesting:

1. No skill to disclose. The positions do not outperform in the first place.
2. Skill decays inside the lag. Real, but stale by 135 days — a finding about how fast edges decay, which is worth knowing.
3. Disclosure is efficiently priced. Thousands of machines parse EDGAR on arrival; there is no drift to capture.
4. Real, and not obtainable — the drift exists and costs consume it. Resolves at the implementation child (ADR-0027), never here.

Falsifier

EV per trade in net R fails to exceed the declared floor, or realised trade frequency falls below the declared minimum, measured over the measurement partition on the rules below and no others.

The control

Funds buy stocks that have been rising. Without a control this measures momentum with a 13F sticker on it.

Each disclosed name is matched to non-disclosed names on trailing return over the disclosure lag, market capitalisation, sector and liquidity, and the hypothesis is about the difference.

The matched control does double duty here, which matters for the independence problem below: it removes the momentum confound and strips out most of the common market factor shared by names entered on the same filing date.

The independence problem — and it is the likely killer

13F filings cluster on the deadline. Entries therefore fall on roughly four dates a year, and positions are held simultaneously with overlapping horizons. Trades entered together are not independent observations.

With k trades sharing a date and cross-sectional correlation rho, the effective count is:

N_eff  =  k / (1 + (k - 1) * rho)

At k = 60 and rho = 0.30, sixty trades supply about three independent observations. At rho = 0.10, about nine.

Over fifteen years at four dates a year — sixty filing dates — effective N lands somewhere near 180 to 540 against the 817 the table below asks for at a 0.15R floor. **That is marginal at best and insufficient at worst, and it turns entirely on rho**, which is measurable on the definition period and must not be assumed.

Naive trade counting would report several thousand trades here and be wrong by an order of magnitude. Matching removes most of the shared market factor, which is the main thing standing between this Draft and arithmetic that does not work — so the control is not a robustness nicety, it is load-bearing twice over.

Search space — declared, enumerated, closed

search_space_size = 8

Any parameter added later re-registers at the larger size.

Power

N = ((z_alpha + z_beta) * sigma / effect)^2. At 8 cells: alpha 0.00625 two-sided, z_alpha ~ 2.73; at 80% power z_beta = 0.84.

Trades required, at a provisional sigma = 1.2R:

These are counts of independent observations, not of trades. Apply the N_eff deflation above before comparing to anything. sigma = 1.2R remains a placeholder to be estimated on the definition period.

Data — and the part that is free

13F filings are free. SEC EDGAR, structured XML for recent history, filings back to the late 1990s, with acceptance timestamps.

And a second-order consequence worth more than the price: a 13F filing is self-point-in-time. It is a dated document stating what was true on a date. For a universe defined as names appearing in 13F filings, the survivorship problem ADR-0025 exists to catch is solved at zero cost — the filings are the constituent history, and no point-in-time universe licence is needed to construct it.

That answers the constituent half of M0.5 for free. It does not answer the rest:

The price series remains the expensive half, and delisted names are exactly where it is dearest — which is also exactly where survivorship bias lives, so it cannot be skipped.

Tier and its child

Mechanism: does disclosed high conviction predict abnormal returns net of the matched control? Carries the large alpha.

Its implementation child, once resolved in favour: does a specific StrategySpec clear its floor net of costs and obtainability? World 4 resolves there.

What would make a null uninformative

• The control is omitted or specified after seeing results — then the verdict cannot separate disclosure drift from momentum.
• **rho is assumed rather than measured**, and effective N is overstated by an order of magnitude.
• Amended filings are used in place of as-filed — a look-ahead aimed directly at the variable under test.
• Delisted names are dropped because their prices were unaffordable, reintroducing survivorship into a design whose universe was survivorship- free for free.