Forensic Signals
The accounting tells, scored from SEC XBRL and backtested out-of-sample. Short the books that look cooked; long the books that are clean. The live screen is members-only.
The validated edge
Measured on ~3,000 US filers, event = the 10-K filing date (no look-ahead). The short cards headline the raw move — how far the stock actually fell, what a short-seller earns — with the market-relative (vs-S&P) figure as a footnote; the long card is judged vs the market, since that's the alternative. The validation gauntlet (held-out, bootstrap CIs, sign-stability, Benjamini-Hochberg FDR) runs on the market-adjusted return. These survived.
Beneish manipulation flag
When the earnings-manipulation model trips, the stock falls over the next three weeks roughly two times out of three.
Cleanest-accruals decile
Companies whose earnings are fully backed by cash flow beat the market by double digits over the following two quarters — compounding to ~+25% at a year. (A long is judged vs the market — that's the alternative.)
Any forensic flag
The broad short basket — any of the four flags — tends to drift lower on a one-month horizon.
The live screen
3 names on the short watch · 4 on the clean-long screen, right now.
Members see every ticker, its forensic score, the exact flag, and the backtested edge per signal — live, the day the filing posts. Free readers see the methodology and the 90-day-delayed reveal below.
What members saw — revealed
Forensic flags go public 90 days after members see them live. Across 130 past short flags, 62 (48%) fell relative to the market since the signal. Here’s the spread — the big wins and the ones that didn’t work, because honest beats cherry-picked.
| Ticker | Flag | Flagged | Since vs SPY |
|---|---|---|---|
| PSIX | High accruals | 2026-03-02 | -76% ✓ (short won) |
| RUN | High accruals | 2026-02-26 | -60.9% ✓ (short won) |
| WING | High accruals | 2026-02-18 | -60% ✓ (short won) |
| UUUU | Inventory outrunning sales | 2026-02-26 | -58.4% ✓ (short won) |
| KTOS | High accruals | 2026-02-23 | -57.6% ✓ (short won) |
| KRMN | High accruals | 2026-04-03 | -55.2% ✓ (short won) |
| ADMA | Beneish manipulation model | 2026-02-25 | -51.4% ✓ (short won) |
| NVAX | High accruals | 2026-02-26 | -42.2% ✓ (short won) |
| MP | Inventory outrunning sales | 2026-02-26 | -39.3% ✓ (short won) |
| OLED | Inventory outrunning sales | 2026-02-19 | -37.8% ✓ (short won) |
| BW | High accruals | 2026-03-16 | -36.1% ✓ (short won) |
| SOUN | High accruals | 2026-03-02 | -36.1% ✓ (short won) |
A short “wins” when the stock fell vs SPY (green). 90-day embargo. Not investment advice.
Across the whole cohort, here's what Analyst + Trader unlock —
- Every trade, every ticker — entry date, entry price, sizing, exit, realized P&L. The audit trail behind the headline number.
- Win rate, average P&L, biggest winner, ugliest loser, capital deployed per ticker. Stop wondering whether the +60% headline is one moonshot or a real pattern.
- Transcript receipts on every row — the exact quote, episode title, timestamp. Verify we didn't take the call out of context.
- Off-the-books table — commentary-only mentions that didn't meet our INITIATE rubric. The "honorable mentions" we deliberately don't count.
- Hot/cold streak analysis — was the edge real, or did wins just cluster? Drawdown stress test — what's the worst stretch you'd have had to hold through?
- BYOB — swap conservative / balanced / let-it-ride methodologies and watch the same host's numbers change under each lens.
- Everything in Analyst, plus:
- Market-state overlay — does this host outperform when markets are calm, when they panic, or both? Per-host breakdown across calm / normal / high-fear tape (so you can see whether the edge holds when it counts, or just in easy markets).
- Alerts when a host opens, trims, or closes a position — email, SMS, WhatsApp, webhook. Stop refreshing the page.
- Compound alert rules — "Chamath initiates a new position AND markets are high-fear AND it's an AI name" → fire. Build the exact setups you want to act on.
- CSV / API export of every ledger — bring it into your own spreadsheet or model.
- MCP server for Claude Desktop / Cursor / Claude Code (coming soon) — pull this data straight into your AI workflow and ask it questions.
- Real-time intraday quotes on open positions — see the host's current book move tick-by-tick during market hours.
Base-rate signals across thousands of companies, not predictions about any one name — ~60% of forensic shorts fall, which means ~40% don’t. Small/mid-cap screen; borrow cost and liquidity apply to the short leg. Educational only. Not investment advice.