US-stock skilled-LLM Agents

Enter a ticker → its latest 10-K runs through Task 2 → an LLM forms a thesis and picks one executable strategy from a fixed menu, grounded in the filing (with citations). A filing-date-aligned backtest then tests it — signals act only after the 10-K was public, so there is no lookahead — shown as a candlestick with entry/exit markers, an equity curve vs buy-and-hold, and honest metrics (losses included). Quarantined 10-Ks are refused, not strategised.

Enter a ticker → its SEC Form 4 filings are fetched and parsed into open-market insider transactions, keyed off each filing's filing date (not the trade date) so the backtest can only act once public. An LLM picks one strategy from a fixed insider-signal menu (cluster buying / net-$ buying / …) grounded in the as-of readings, then a lookahead-free backtest runs vs buy-and-hold and the S&P 500. Only open-market buys/sales count — grants and option exercises are excluded, and selling is a weak exit signal, not a short.

Enter a ticker → we resolve its sector ETF from the SEC SIC code (S&P 500 fallback) and compute a relative-strength series (stock ÷ sector), strictly as-of the latest close. An LLM picks one strategy from a fixed RS menu (uptrend / breakout / momentum); a lookahead-free backtest then holds the stock long/flat with RS deciding only when to be long — vs buy-and-hold and the S&P 500. Surfaces nuances like “beating the market but lagging its own sector.”

Enter a watchlist → each name's long/flat signal comes from the Task 4 agent; an LLM picks one sizing policy (equal-weight / inverse-vol / risk-parity / signal-proportional + single-name cap, gross cap, vol target, rebalance) from as-of universe stats — not per-name weights, which are deterministic. A lookahead-free portfolio backtest then allocates across names vs an equal-weight basket and the S&P 500. The piece that turns the signal agents into a sized, risk-controlled book.

Enter a ticker → its recent earnings press releases (8-K Item 2.02 / Ex-99.1) are fetched; an LLM classifies each as-of its filing date (sentiment / guidance / beat-miss, with citations); then it trades the post-earnings drift — lookahead-free, acting only on the open after each filing, vs buy-and-hold and the S&P 500. Reads the press release, not the live Q&A transcript (source is pluggable).

Enter a ticker → we track a curated set of well-known managers (Berkshire, Baupost, Pershing Square, …) via their SEC 13F-HR filings and follow whether they're accumulating the name. An LLM picks a follow-the- smart-money strategy; the backtest is keyed off the 13F filing date (~45-day lag → a lookahead-safe but slow confirmation signal). Curated funds only, matched by issuer name — both surfaced honestly.

Splits returns into the overnight (close→open) vs intraday (open→close) move — the documented anomaly that most US-equity return accrues overnight — then backtests a participation rule honestly net of the daily round-trip cost that usually erases the gross edge.

FINRA daily short-volume ratio (weekly-sampled, cached) → a squeeze / low-short rule. Honestly flagged: this is short volume (incl. market-maker hedging), not short interest — free historical short-interest doesn't exist, so it's the closest free proxy.

Three classic free quality factors from SEC XBRL — Piotroski F-Score, Sloan accruals(earnings quality), and the asset-growth anomaly — point-in-time, filing-date keyed; the LLM picks the factor or a composite. Numbers, where T3 reads the 10-K text.

Schedule 13D activist stakes (positive drift) + red flags (dilution, late filings, auditor changes, delisting, adverse 8-K 5.02 exec departures the LLM reads from the text). Keyed off the filing date; ride activist drift, stand aside on red flags.

Give it a watchlist; the LLM picks one long-only cross-sectional factor (12-1 momentum, low-vol, near-52w-high, or short-term reversal). Each rebalance the universe is ranked on trailing data and the top-N is held — benchmarked against the equal-weight basket of all names and the S&P 500.

A bellwether's earnings move its peers before they report. Classify the bellwether's 8-Ks (reusing Task 8) and trade the peer in the short read-across window — keyed to the bellwether's filing date, lookahead-free.