DRAFT-001 — Forced selling at index deletion

A Draft, in the sense CONTEXT.md gives the word: a proposed Hypothesis carrying a mechanism and a falsifier, with no standing until a human confirms its registration, which is the act that spends alpha. Nothing here has been measured. Nothing here has cost anything.

It exists to make one thing concrete: what a hypothesis looks like when the search space is small because the mechanism is sharp, rather than small because someone trimmed it to lower the correction.

Read this first — it cannot be registered today

The information axis is cross_sectional. ADR-0028 allocated that axis 0, and ADR-0017 is unambiguous about what a zero-budget axis accepts. This Draft is therefore evidence for the amendment note appended to ADR-0028 on 2026-09-06 — the axis set to zero is the one where the retail structural advantage lives — and it is not a request to register anything.

It becomes registrable only if M0.19 finds the narrow event data affordably, and M0.14 leaves room to move the axis off 0.

Mechanism

Funds tracking an index must sell a deleted constituent on or about the effective date. That selling is price-insensitive: it is not a view, it is a mandate, and it happens whatever the price.

If stocks were perfect substitutes, this would not matter — arbitrage capital would absorb the supply at fair value and the demand curve for any single name would be flat. The claim under test is that it is not flat (Shleifer 1986; Harris and Gurel 1986): price-insensitive supply meeting a finite pool of absorbing capital pushes the price temporarily below fair value, and whoever provides that liquidity is paid for it as the price recovers.

The prediction follows directly and has a sign: deleted names are pushed down into the effective date and recover afterwards. The trade is to buy once the forced supply is exhausted and hold through the recovery.

Two properties make this worth asking rather than the version it replaces ("detect institutional accumulation on a chart"):

• The forcing is disclosed, not inferred. The index provider publishes which names, and when. There is no proxy standing between the hypothesis and the event, so a null verdict means something.
• Nobody is hiding it. Execution algorithms exist to conceal discretionary accumulation. An index fund's mandate cannot be concealed — it is published in advance, by design.

Both interpretations

If the effect is real. Demand curves for individual equities slope downward at the horizons that matter, and the compensation for absorbing forced supply survives in names too small for the arbitrage capital that would otherwise flatten it. The effect should be larger where index-fund ownership is a larger fraction of float, and where the absorbing capital is thinnest — a directional prediction, stated here, testable as a pre-declared subgroup rather than discovered afterwards.

If it is not. Three distinct worlds, and they are not equally interesting:

1. Demand curves are flat. Substitutes are close enough that arbitrage equalises price. The mechanism is wrong.
2. It existed and has been arbitraged away. Both source papers are from 1986 and the effect is documented as having decayed. This is the modal outcome and it is a real result — a dated effect is a finding about markets, not a failure of the apparatus.
3. It is real and not obtainable. The raw effect survives but the spread on names small enough to exhibit it consumes all of it. This is the most likely "false" and the most important to separate, because it is a verdict about your cost structure rather than about the world — and it is exactly what the mechanism/implementation split in ADR-0027 exists to keep apart. A mechanism verdict must not be allowed to absorb it.

Falsifier

Measured over the measurement partition, on the entry and exit rules declared below and no others:

EV per trade in net R fails to exceed the declared floor, or realised trade frequency falls below the declared minimum.

Either half alone falsifies. The floor is a pair and both halves bind (CONTEXT: Detectable floor).

The control that makes this a real hypothesis

Deletions are not random. A stock is deleted because it fell. Buying deletions is therefore, by construction, also a bet on reversal in beaten-down small caps — a well-known and heavily-mined effect that has nothing to do with index flow.

Without a control, a positive result here is uninterpretable: it would be consistent with the mechanism being entirely false and the returns coming from a factor already in the literature.

Declared before running, not after: each deleted name is matched to non-deleted names in the same universe on trailing drawdown, market capitalisation and liquidity over the same window. The hypothesis is about the difference, not the raw return. If matched non-deleted names show the same recovery, the mechanism is false however profitable the raw trade looks.

This is the single most important line in this document. A version of this hypothesis without the matched control is a repackaged reversal factor wearing an index-flow costume, and it would spend mechanism alpha to learn nothing.

Declared exclusions

• Deletions caused by merger or acquisition are excluded. The price is pinned to deal terms, there is no liquidity vacuum, and the return distribution is a different object entirely. Requires reason codes or a corporate-actions series (M0.4).
• Deletions caused by delisting or insolvency are excluded on the same grounds and with the same dependency.
• Only deletions at scheduled reviews are in scope. Ad-hoc intra-review removals have a different announcement-to-effective interval.

Search space — declared, enumerated, and closed

search_space_size = 4

Index families are pooled into a single test, not tested separately. Testing them one at a time multiplies the search space by the number of families and buys nothing the pooled test does not already answer.

Any parameter added later re-registers the hypothesis at the larger search space. The correction is a measurement of how much was looked at, not a fee to be minimised — and four cells is honest here only because the mechanism admits four. It is not a trimmed version of a larger sweep.

Power

Per the corrected M0.15 formula, N = ((z_alpha + z_beta) * sigma / effect)^2. At search_space_size = 4: alpha 0.0125 two-sided, z_alpha ~ 2.50; at 80% power, z_beta = 0.84.

Trades required, at a provisional sigma = 1.2R:

**sigma = 1.2R is a placeholder and must not be used.** M0.15 now requires sigma to be a declared parameter with recorded provenance, and this hypothesis is a case where assuming it would matter: a wide stop on a name that has just been dumped produces a fat-tailed outcome distribution, and sigma above 1.5R would move every row in that table by more than a third. It is estimated on the definition period, which is what the definition period is for — never on the measurement partition, which would consume the observations the hypothesis is about.

Trades available is deletions_per_year x years_in_measurement_partition, after the exclusions above remove the M&A and insolvency cases — which may be a large fraction. Both terms are M0 lookups. The arithmetic runs in minutes once M0.19 supplies the first.

Detectable floor — derived, then the author's

Not chosen freely. The lower bound is set by cost:

floor  >  cost_in_R  =  c / s

where c is round-trip cost as a fraction of notional and s is stop distance. M0.16 produces the admissible band, M0.2 produces c. On small-capitalisation names c is large, so a viable s is wide, and a wide s interacts with the stop question below.

The author sets the floor within that band and declares the minimum trade frequency alongside it, before anything runs. The number is not recommended here and no default appears anywhere.

Required capabilities

Small-capitalisation cash equity at the venue · share granularity fine enough that a minimum-size position is possible in a low-priced name · a native stop · sufficient depth that a minimum-size order does not itself move the price.

A live tension for the implementation child. The StrategySpec mandates a stop, and a stop placed on a name that has just been force-sold into a liquidity vacuum is standing exactly where the volatility is. If the stop is tight it will be hit by the tail of the very selling the hypothesis is about; if it is wide, cost_in_R falls but position size falls with it. This is an implementation question, not a mechanism one, and it belongs to the child.

Tier and its child

Mechanism hypothesis (ADR-0027): does the effect exist, net of the matched control? Carries the large alpha.

Its implementation child, which cannot exist until this resolves in favour: does a specific StrategySpec — one entry, one exit, one stop, one sizing rule — clear its floor net of costs and obtainability? World 3 above resolves there, not here.

What must close first

What would make a null uninformative

Recorded now, because a null costs mechanism alpha and counts permanently against the lab falsifier (ADR-0033), and a null that teaches nothing is the most expensive outcome available.

• The matched control is omitted or specified after seeing results. Then neither verdict distinguishes index flow from ordinary reversal.
• **sigma is assumed rather than measured**, and N is set from a number that flatters the design.
• The exclusions are not executed because the reason codes were unaffordable, leaving acquisition arbitrage mixed into the sample.
• The universe is not point-in-time (ADR-0025). Reconstructing "which names were in the index then" from today's membership is a survivorship error aimed directly at the variable under test.

If any of these holds at registration time, the honest action is to refuse the registration rather than spend the alpha.

Correction, 2026-09-10 — the independence problem applies here too.

Found while drafting DRAFT-002 and it belongs here as much as there. Scheduled index reviews are quarterly, so deletions cluster on roughly four effective dates a year and the resulting positions are held simultaneously over overlapping horizons. Trades entered together are not independent observations, and the power table above counts trades as though they were.

With k trades sharing a date and cross-sectional correlation rho, the effective count is N_eff = k / (1 + (k - 1) * rho). At k = 30 and rho = 0.30, thirty trades supply about three independent observations.

Every row of the power table must therefore be read as independent observations required, not trades required, and deletions_per_year from M0.19 must be deflated by a rho measured on the definition period before M0.15's arithmetic means anything. Naive trade counting overstates power here by roughly an order of magnitude.

The matched control partly rescues this. By measuring the difference against matched non-deleted names it strips out most of the common market factor that drives rho, so the residual correlation after matching is far below the raw figure. That makes the control load-bearing twice over — once against the reversal confound, once against this — and it makes measuring the post-matching rho the number that actually decides whether this hypothesis is affordable.