Library · Markets · Signal Desk
Latest nightly tapeThe latest tape, misses and all.
At 00:30 PT the desk tested 2,488 operating-company shares. A setup had to beat that stock’s ordinary five-day result after costs, survive a check on newer data, and be operating common equity by its registered name. Zero is an acceptable answer.
Trading date 2026-08-24
generated
- stage 17,130symbols with 350+ daily barsthe archive-backed sweep universe before any gate
- stage 24,642excluded up frontobvious non-company classes, thin filing records and non-common share lines; known fund/SPAC leak disclosed below
- stage 32,488companies tested5 setups · older data for research, newer data held back as a check
- stage 40made it through the maththe plausible range cleared the normal result after costs and the multiple-test penalty
- stage 50left on the watchlistno candidate was excluded by an instrument rule
Nothing left on the watchlist.
That is a real result, not an empty-state error. With 5 setups, trading costs and a penalty for trying several ideas, the desk is allowed to say “nothing.” We publish that answer because a track record made only from interesting nights is not a track record.
How a name gets onto the tape
01 · Start with companies
The desk starts with US-listed common shares whose companies file periodic SEC reports and have enough fundamentals to score. That removes the obvious funds, preferreds, warrants and units. A small group of 10-K-filing funds and seasoned SPACs can still leak through; that limitation is measured, not hand-waved.
02 · Make it pay its costs
An up move does not automatically count. It has to clear a deliberately pessimistic round-trip cost — as much as 1% for cheap shares. A paper edge that disappears after the spread and slippage is not an edge.
03 · Beat normal, with room to spare
A 70% backtest can be meaningless if the stock rose 72% of the time anyway. We compare every setup with its normal result, put a plausible range around the estimate, and raise the bar for every factor we have tried. Cherry-picking the winner is not allowed.
04 · Check newer, unseen data
The setup is built on an older window, then checked against a newer slice it did not see. That smaller slice is not asked to prove the idea again. It is there to catch a story that immediately falls apart out of sample.
05 · Exclude by rule, not by opinion
A fixed name-pattern rule excludes funds, trusts, SPACs and derivative share classes that survived the universe gate. The rule that fired and the text that triggered it are both published, so any exclusion can be reproduced. A local model used to write this judgement in prose; it was measured leaking post-date outcomes on 4 of 4 distressed probes, so it no longer decides anything and is kept only as an audited check on what the rules miss.
→ · Then let paper trading argue back
A surviving name becomes a market-on-open paper order with a 5-day exit. The desk compares real fills with the backtest every week. At 30 closed trades it produces the first decay report. A setup that stops working gets retired in public.
Ranked long candidates from a walk-forward factor screen over IEX EOD bars (~2.5% of consolidated volume). Universe: US-listed common shares of SEC periodic filers with scoreable fundamentals — this removes closed-end funds, index ETFs, thin-XBRL trusts, and preferred/unit/warrant classes up front, but some 10-K-filing funds (commodity/crypto pools, BDCs) and seasoned SPACs can remain; a rule-based instrument classifier flags those from the registered name and its exclusions are published. Hit rates carry Wilson intervals and Bonferroni correction; thresholds model round-trip costs. Educational research, not investment advice.