Research, not recommendations
Scan a broad U.S. stock universe for sessions trading more than 9× normal volume, compare each one with earlier look-alike events, and check whether the pattern held on later data.
Event-day volume vs. the prior 60 sessions, split-normalized, event day excluded.
Explainable nearest neighbours using only outcomes already known at the signal.
Walk-forward tests, a held-out period, costs, and clustered uncertainty.