{"api_version":"v1","endpoint":"what-we-tested","one_sentence":"I tried four times to prove otherwise, and four times the data said no.","summary":"Four pre-registered, sealed tests for a forward edge in SEC Form 4 insider buying. All four KILL -> NULL: no forward edge distinguishable from the same-week same-universe base rate on our in-sample window. A KILL means 'no edge demonstrable here', not 'no effect exists'. We sell filing-verified facts, not a predictive signal.","tests":[{"id":1,"question":"Do insider cluster buys (multiple insiders buying the same issuer in a window) beat the base rate?","method":"Entry on filed_date (public), control = equal-weight mean forward return of all priced purchase events in the same ISO week; block-bootstrap (block=5, 10k resamples), BH-FDR q=0.10; PASS needs edge >= +1.5%, significant, breadth >= 55%.","sample":"Study 1: 151 U.S. trading days; 1,385 priced (ticker, filed_date) purchase events.","number":"+10d primary edge +0.283% (p=0.730), breadth 46.5%.","verdict":"KILL - fails all three legs (edge, significance, breadth).","source":"docs/research/CORPUS_SIGNAL_RESULTS.md"},{"id":2,"question":"Do large >=10%-of-holdings insider buys beat the base rate?","method":"Same frozen harness, control and PASS gate as test 1 (same-week same-universe control, BH-FDR q=0.10).","sample":"Study 1: same 151-day window, 1,385 priced events.","number":"+10d primary edge +1.147% (p=0.332), breadth 47.4%.","verdict":"KILL - fails all three legs.","source":"docs/research/CORPUS_SIGNAL_RESULTS.md"},{"id":3,"question":"Do large officer buys in small-cap issuers beat the base rate?","method":"Study 2 (CMP officer/size conditioning); officer flag OR'd per event; small-cap = bottom cap tercile AND cap < $2.0B (cutoff $306M); +21d theory-matched primary horizon; same control + block-bootstrap + BH-FDR q=0.10 machinery.","sample":"Study 2: 157-day window; n=116 priced small-cap officer events at +21d.","number":"+21d edge +9.74% (p=0.243, 95% CI [-3.64%, +27.89%]), breadth 43.1%.","verdict":"KILL - genuinely powered (n>=100) but not significant and breadth < 55%; the large point estimate is carried by a minority of names.","source":"docs/research/CORPUS_SIGNAL_2_RESULTS.md"},{"id":4,"question":"Do opportunistic (vs routine-cadence) officer buys beat the base rate?","method":"Study 2; routine-proxy = buyer with >=3 distinct buy-months spanning >=90 days removed, rest opportunistic; +21d primary; same frozen inference.","sample":"Study 2: n=468 priced opportunistic officer events at +21d (only 18 routine-proxy buyers existed to remove).","number":"+21d edge +4.21% (p=0.157, 95% CI [-0.98%, +10.92%]), breadth 44.7%.","verdict":"KILL - not significant, breadth < 55%; removing routine buyers did not sharpen the signal.","source":"docs/research/CORPUS_SIGNAL_2_RESULTS.md"}],"caveat":"8-month in-sample window on 2026 tape, small samples, Study 1's 151-day vs Study 2's 157-day windows. Study 2's routine/opportunistic split is the weakest limb: only 18 routine-proxy buyers, far short of the CMP literature's >=3-year cadence requirement. A KILL means 'no edge demonstrable here', not 'no effect exists' - a positive-but-insignificant point estimate cannot be separated from noise at this power.","rerun":"All four tests will be re-run under the same sealed pre-registrations once the Form 4 backfill covers two years, with the result published either way."}