heurist-finance: Skill for Claude Code

Technical Evidence

Shared plumbing and output contract: ../_shared/analyst-core.md. TA's honest job in this stack: timing and risk placement for a thesis that exists on other grounds — and occasionally a standalone signal where the effect is actually documented. (For the operational entry/exit procedure — regime-gated band reads, limit-price ladders, stops — use entry-exit-timing; this skill is the evidence layer beneath it.) This skill's core content is the evidence tiering, because treating folklore and documented effects as one "technical picture" is the signature TA failure.

The evidence tiers

Tier 1 — documented, still-live effects (usable as signal):

• Short-term reversal: ~1-week losers beat winners over the following week (Jegadeesh 1990) — liquid names, fades with size. Implication: don't chase multi-day spikes; buy pullbacks, fade panic days in quality names.
• 52-week-high proximity (George & Hwang): anchoring underreaction near highs, ~0.5%/month effect. New-high lists are a feature, not a warning.
• Momentum 3–12 months — full treatment in momentum-trend; it's the strongest price-only effect known.
• Post-earnings drift — price/volume confirmation of a surprise carries real forward signal (earnings-play).
• Volatility clustering: today's vol predicts tomorrow's (the one universally replicated price fact) — position sizing and stop width should be ATR/realized-vol-scaled, always.

Tier 2 — degraded or conditional (context only, never standalone):

• Moving-average trend rules: genuine in-sample edge on indices pre-1990 (Brock, Lakonishok & LeBaron 1992), largely arbitraged away after publication and costs. Still useful as regime descriptors (above/below 200d changes the return distribution's shape) — not as entry signals.
• RSI(2)-style short-term mean reversion (Connors): worked in specific ETF samples, regime-dependent, dies in trends. Test on the actual asset's recent data before citing (see below).
• Volume confirmation: breakouts on ≥1.5× average volume fail less often — directionally supported, weak alone.
• Classic patterns (head-and-shoulders etc.): mixed academic evidence (Lo, Mamaysky & Wang found some information content; out-of-sample profitability after costs is thin). Descriptive language, not signal.