Sweeper — volume cohorts

Outcome distribution — what to expect, where to stop

The cohort tables below show which alerts work; this one shows the distribution of what happens after any alert — every number is the day-N close (a price you could transact at), measured from the intraday alert price. Volume alerts fire mid-session, so day 0 is already a held position. close return percentiles p10 / p25 / median / p75 / p90 give the spread of outcomes; the median is the typical case. dd-p10 is the deep-drawdown tail over day 0..N — a stop has to sit below dd-p10 to survive the typical-bad alert. green = share closing positive. edge vs SPY = median of (alert return − SPY return) over the same day-0 → day-N window, measured close-to-close — a positive median is only a real edge if this is also positive; if it sits near zero the alerts are just riding the market. (A bracketed count shows how many alerts had SPY data when it is fewer than n.) 598 alerts; n shrinks down the table as fewer alerts are that mature.

Cohorts — which alerts work best

For each cohort, three maturity windows side by side: day 0 (alert-day only), day 1 (alert day + next session — only counts alerts that have a next-day bar), settled (only alerts with all 5 days complete). Comparing them tells you whether alerts hold up over time or fade. Empty rows = no alerts in that bucket yet. "hit@+2%" = % of alerts in the bucket that reached at least +2% high during the window.

By sector

By RVAT bucket

By time of day

By weekday