Loved this piece. I ran your two ideas against a 20-year Form 4 cluster signal (~5,000 observations) and the numbers line up beautifully.

Your 80%-same-date-and-price filter is the highest-leverage piece in the article — applied to my baseline panel it rejected 3,682 of 4,912 of my "clusters" as ESPP/directed-share noise (75%), with max_dup_share of exactly 1.0 in many cases (every contributing transaction identical date and identical price). The Cohen-Malloy-Pomorski routine filter catches only a tiny fraction of these. Your filter is doing genuinely different work.

Your Ozlen-Batumoglu timing claim translates cleanly to portfolio mechanics: on the cleaned cluster set (n=688 trades, 21-day hold, −5% stop, 50 bps round-trip), filing+1 morning entry gives Sharpe 0.65 OOS; filing-day intraday VWAP gives 1.76; transaction-date entry (the theoretical ceiling) gives 2.17. Disclosure-delay tax of ~1.5 Sharpe points; intraday filing-day execution recovers ~80% of the ceiling — almost exactly your 70–80% figure for the unobservable-window share.

Director and 10%-owner lanes generalize at full sample with similar stories. Happy to share details if useful.

Thanks for reopening a lane I'd archived as a kill.