Filings didn’t predict share prices. So I use them to predict the next set of numbers.
First test: across 1,046 filings (2019–26), simple margin and growth signals called the next 3 months’ share-price move right 49% of the time, no better than chance. So the question changed. Each filing now runs through five investor lenses and ends in specific calls on the next filing’s numbers, each with a confidence level, written down before the results. Every call is scored, misses included.
- COMPANIES ANALYSED11
- PRE-REGISTERED CALLS53
- RESOLVED SO FAR4 of 5 right
- TO SCORE BY 5 DEC30 calls
CALL LEDGER · SAMPLE
CONFIDENCE · RESULT- Tesla · Q3 regulatory credits below $500m65%Right: ✓ $113m
- Tesla · Q3 free cash flow positive60%Right: ✓ +$1.5bn
- Tesla · Q3 energy margin above auto ex-credits80%Right: ✓ 33% vs 18%
- Tesla · nine-month capex at least $13.0bn75%Right: ✓ $13.01bn
- Tesla · Q3 gross margin between 16% and 19%70%Wrong: ✗ 19.6%
- Nvidia · Q3 FY27 revenue at least $105bn60%Open, due 5 Dec
- Supermicro · material weakness still reported80%Open, due 30 Nov
THE LENS STACK
EVERY FILING- ForensicCan the numbers be trusted?
- BuffettIs this a durable business?
- MarksWhat is already priced in?
- SituationWhat matters in this setup?
- Red teamWhat would break the call?
HOW IT WORKS
PIPELINE- 10-K / 10-Q
- Evidence record
- Forensic overlay
- Buffett · Marks
- Red team
- Scored calls
Only 5 of 53 calls have resolved (all Tesla, Q3 ’26): too few to judge skill; about 30 are needed. Calibration so far: Brier 0.175, against 0.25 for always saying 50%. The backtest universe was chosen in 2026, so it carries survivorship bias. No positions are taken.
WHAT’S NEXT
ROADMAP- Score each call as results are published: 30 scored by 5 Dec
- Backtest v2: test language change as a signal, with delisted companies added
- Link Filing Words: feed risk-wording shifts into each call