Public track record
Every high-conviction thesis, graded.
We publish forward returns for every Form 4 AI thesis at conviction ≥70, benchmarked against SPY over the same window. Updated nightly. No cherry-picking — everything is here, including the calls that didn't work.
No settled high-conviction calls yet
First entries appear after the 7-day settlement window.
We refuse to grade trades before there's been time for them to play out. As soon as a filing scored ≥70 crosses 7 days post-filing, its forward return vs SPY lands here automatically — no editorial selection.
Methodology
How we score the calls.
- Universe: every AI thesis we generate at conviction ≥70. No discretionary selection. No survivorship bias. Synthetic / test-data filings are excluded.
- Settlement window: we don't grade until a filing is at least 7 days past its EDGAR publication date. This gives the market time to absorb the information and avoids reporting noise that hasn't yet resolved.
- Return computation: for each filing we compute close-to-close return on the issuer's ticker over 7 / 30 / 90 days from filed_at, then subtract SPY's return over the same window. Both pulled from Polygon. Days where either ticker doesn't trade roll forward to the next session close.
- Direction handling: the conviction score is direction-aware (a high-conviction sell is a bearish call). We sign returns accordingly so a +5% move on a buy and a −5% move on a sell are both shown as a +5% "hit."
- Sample size warning: insider-buy backtests are notoriously noisy. With fewer than ~30 settled calls, treat aggregate numbers as directional only. We publish raw counts, not just averages, so you can apply your own significance bar.
- Updated: page revalidates hourly; price data is reread once a day. The full dataset is downloadable from /filings.
Reminder
This is journalism over public regulatory filings, not investment advice. Forward-return tables are noisy at small sample sizes; treat early numbers with skepticism. Past insider performance does not guarantee anything; insider-buy backtests are notoriously path-dependent.