Setup: Buyback-announcement drift

ID: S020 Slug: buyback-announcement-drift-v1 Type: EVT Date added: 2026-06-08 Status: open (data scraper required) — promoted from lab/candidates/buyback-announcement-drift-v1.md Wall: 3 — Messy data that won't license cleanly

Data infra dependency (partially resolved 2026-06-13): the PIT shares-outstanding layer now exists (src/data/edgar_facts.py, SEC companyfacts, ~94% coverage) — built for S017, reused here. That unblocks the realized-buyback angle below (observing share-count declines) immediately. The original announcement leg still needs the SEC EDGAR 8-K scraper + item-type classifier (shared with S019); that part of Stage 1 waits on the scraper.

One-line setup

We expect that small-cap names that file an 8-K announcing a share repurchase programme materially large relative to market cap, entered shortly after the filing posts and held over a multi-week drift window predict positive mean excess return over an equal-weighted small-cap baseline in the US small-cap universe with a liquidity floor over a 20-60 trading-day forward window, because a buyback is management spending real money on the view that the shares are cheap, and it mechanically shrinks the share count, lifting per- share value; the market tends to react gradually rather than instantly, leaving a documented post-announcement drift that's larger and slower in smaller names with less analyst coverage.

Rationale (the "because", expanded)

A buyback is management spending real money on the view that the shares are cheap, and it mechanically shrinks the share count, lifting per-share value. The market tends to react gradually rather than instantly, leaving a documented post-announcement drift — and the drift is larger and slower in smaller names that get less analyst coverage.

Shares the EDGAR-scraping plumbing with S019 (insider-cluster-buys). Orthogonal to the vol/trend book; calendar-mechanical but event- triggered rather than date-triggered.

Data required

• SEC EDGAR 8-K filings with item-type classification (the scraper shared with S019) — keying on filing type + a few repurchase-announcement phrases, not free-text NLP
• Tiingo EOD prices + market-cap data
• Universe filter: small-cap band + liquidity floor
• Filter on programme size relative to market cap (token buybacks excluded — calibrate threshold)

Quick-kill gate (Stage 1)

Will be considered to have passed Stage 1 if:

• Mean per-event 30-day excess return over equal-weighted small-cap baseline ≥ +1.0pp [suggested, to freeze]
• Hit rate ≥ 53% [suggested, to freeze]
• Sample size ≥ 400 announcement events [suggested, to freeze]
• Effect present in both sample halves [suggested, to freeze]
• Welch p<0.05 AND mean>0 [suggested, to freeze]
• Size-conditional check: events where the announced programme is ≥ 5% of market cap [suggested, to freeze] must outperform smaller-programme events — confirms the "real signal vs token" distinction matters

Realized-buyback confirmation (pre-registered angle, added 2026-06-13)

Announcing a buyback ≠ executing one — the spec already flags this as the main risk. Now that PIT shares outstanding is available (edgar_facts), we can condition the drift on execution evidence: does the next 10-Q/10-K actually show shares outstanding declining?

• Signal: for each announcement, check whether trailing shares outstanding fell by ≥ [X]% over the following 1–2 filings.
• Hypothesis: the executed subset drifts more cleanly than the announcement-only set; token/never-completed programmes are the noise the announcement signal carries.
• Status: pre-registered exploratory (non-gating) for v1 — it cannot pass or fail S020 by itself. If it is the real driver, it becomes the core of buyback-execution-drift-v2 under a new S-number, NOT bolted onto v1 mid-test (per lab discipline).
• Runnable now on the shares side (no 8-K scraper needed): a pure realized-share-count-decline drift can even be tested independently of the announcement leg as an early read.

What I expect to find

Effect probably present but smaller and noisier than the insider- buys signal — announcing a buyback is not the same as executing one, and many programmes are never completed. Expected mean excess return 0.5-1.5pp, hit rate 51-55%. Probability of clearing the gate is moderate (~45%). Most likely failure mode is the friction-sensitivity gap: in the smallest small-caps where the effect is strongest, the round-trip spread can eat 100-200 bps and leave only a marginal edge.

Notes

• Lower priority than S019: noisier signal, same EDGAR cost. Run after S019 to amortise the scraper investment.
• Announcing ≠ executing. Some programmes get cancelled or under- executed. The size-conditional gate exists to favour announcements with skin in the game (large-relative-to-cap programmes).
• Look-ahead trap: 8-K filing timestamp is the entry trigger. The announcement date in press releases occasionally precedes the SEC filing by hours; use the filing timestamp to be safe.

Disclosure boundary

This setup file is internal. Downstream result.md / kill.md writeups must follow lab/DISCLOSURE_POLICY.md §2. Pre-publish: python -m pytest tests/test_disclosure.py.