A faster horizon: what happens around the filing itself

This one survives — with caveats. Windows C and D (reaction from the filing’s close through the next session) show a strong, consistent gap between buy days and sell days (t ≈ 8), the opposite of the null monthly-panel result. It holds with the single most-concentrated ticker removed, it grows rather than reverts over the following ten trading days (day+1 0.63% → day+10 0.90% — a bid-ask bounce would fade, not build), and it is not an earnings-date artifact (excluding filings within three days of a 10-Q/10-K leaves the spread unchanged). The likely reason the monthly panel misses this: it nets buys against sells into one continuous score, and insider buying, not selling, is the informative side in the literature — selling has too many routine reasons (diversification, taxes, scheduled 10b5-1 plans) to carry a clean signal once averaged over a month. Isolating pure buy days from pure sell days at a short horizon appears to recover what the monthly netting buries. Not yet accounted for: real transaction costs and bid-ask spreads on names this illiquid could erase much of a headline gross return this size, and no size or sector controls have been applied to this design.