Stats
Actions
Available In
Tags
quant-llm-skills
🇺🇸 English | 🇪🇸 Español | 🇨🇳 简体中文
Skills for quant research with LLMs that don't fall for the traps nobody talks about.
Most "AI trading" skill packs teach Claude how to backtest CANSLIM and parse 10-Ks for AAPL. They fail the moment you point them at a real small-cap with sparse XBRL, an active ATM, and four 13D filers reporting the same shares.
🎯 Especially valuable on Haiku and smaller models. Our regression suite shows 8–9 out of 11 evals measurably differentiate Haiku output — caught only with the plugin loaded. Cost-conscious users running cheaper models are silently shipping broken analyses. Reproduce in ~100 seconds:
make evals-baseline. See BENCHMARKS.md.
This pack distills hard rules from running production quant pipelines into Claude Code skills the LLM applies automatically — no system-prompt hacking, no manual invocation, no extra glue.
What it catches that a baseline LLM misses
- Lookahead bias in any historical query —
period_endis not a publication date. - Capacity vs action — an effective S-3 alone is not a dilution event.
- ATM vs registered direct vs ELOC — same 424B5, different trades.
- The placement agent matters — Goldman on a secondary ≠ Wainwright on an ATM.
- XBRL gaps for FPIs and SPACs — the data is in the filing text, not the structured feed.
- 13D group double-counting — naive sum of cover pages overstates insider ownership 2–10x.
- Quantitative dilution scoring — 0–100 with auditable component weights, not vibes.
See EXAMPLES.md for real prompt-and-response transcripts.
Skills
| Skill | What it does |
|---|---|
| lookahead-safety | Forces filing_date as known-date, never period_end. The #1 quant-backtest bug. |
| sec-filing-types | Disambiguates SEC forms (S-3, 424B, 8-K items, 13D/G, Form 4, 20-F, 6-K, NT 10-K). Knows that a shelf is capacity, not action. |
| atm-detection | Multi-signal inference for active ATMs. Distinguishes ATM from ELOC and registered direct. Catches dilution that 8-K-only scanners miss. |
| bank-tier-classification | 4-tier framework mapping placement agents (bulge bracket → small-cap specialist) to expected deal behavior. |
| xbrl-fallbacks | When SEC XBRL is empty or 404 (FPIs, recent IPOs, SPACs), defines the cover-page hierarchy and extraction rules. |
| dilution-event-scoring | 0–100 framework integrating ATM + agent tier + recency + cash runway + structure + history. Reproducible, auditable, with action thresholds. |
| insider-dedup | Joint-filer / group / family-attribution dedup rules for 13D/G and Form 4 aggregation. Stops the cover-page-sum bug. |
| code-review-for-quant | Domain-specific code-review checklist (lookahead, splits, snapshots, NaN propagation, joint-filer dedup). Ranks bugs by silent-corruption potential, not by severity-of-symptom. |
| transaction-cost-modeling | Realistic friction defaults for small caps. Catches borrow APR fiction (3% on Reg-SHO names instead of 50–500%), locate-failure-as-slippage bugs, and engine-default near-zero friction. |
| survivorship-bias | Catches the universe-built-from-today's-survivors trap. Special focus on small-cap patterns: reverse-split-then-delist phantom returns, ATM-into-delisting, SPAC merger flips. |
The skills compose: ask "score X's dilution risk" and the scoring skill calls the ATM, agent-tier, and lookahead skills automatically.
Install
Once published to GitHub:
/plugin marketplace add jefrnc/quant-llm-skills
/plugin install quant-llm-skills@quant-llm-skills
For local testing now:
/plugin marketplace add /absolute/path/to/quant-llm-skills
/plugin install quant-llm-skills@quant-llm-skills
Or one-shot via CLI without installing:
claude --plugin-dir /absolute/path/to/quant-llm-skills -p "your prompt"
