heurist-finance: Skill for Claude Code

Portfolio Review

Shared plumbing and output contract: ../_shared/analyst-core.md. Positions are sized one at a time (position-sizing), but ruin happens at the book level — usually because ten "different" positions turn out to be one bet. This skill audits the book the way a risk desk would: not "are the picks good?" but "what is this portfolio actually one big bet on, and what happens when that bet has its worst month?"

Input: the user's holdings (tickers, sizes, cost basis if offered, open options positions with strikes/expiries). Ask for them if not provided — there is no useful portfolio review without the actual book.

The audit, in order of how often it finds the problem

1. Effective number of bets (the headline finding, usually). Cluster holdings by shared driver, not by sector label: four AI-adjacent names + QQQ + a semis ETF = one trade with six tickers. Method: pull 1–2y of prices for all holdings (prices ... --out FILE), compute the pairwise correlation matrix in Python, cluster at |ρ| > ~0.6, and — because calm-period correlation understates it — re-cluster using each name's behavior in the last genuine stress window (2022 for rates, 2020 for growth-shock). Look-through ETFs with fund_snapshot before clustering: overlapping top holdings across "diversified" ETFs is the most common hidden concentration. Report: N tickers, ~K effective bets, largest cluster = X% of book.
2. Concentration + liquidity. Top position and top cluster as % of book (top single > ~15% or top cluster > ~35% deserves an explicit conviction justification, not a drift explanation — most concentration is unrebalanced winners, which is a decision being made by default). Days-to-exit per position at ~20% of ADV; a book that takes two weeks to unwind owns a risk the P&L doesn't show (small-cap-scout math, book-wide).
3. Factor and macro tilts. Estimate book beta (position-weighted, from the price file); eyeball the style skew (growth/value, size, momentum-crowding overlap) and translate to macro exposure in plain terms: "this book is long duration + long AI capex + short energy — it has a rates view whether you meant one or not." Cross-check against the user's stated views — the coherence test at book level: a user predicting a correction while running a 1.3-beta, zero-cash book is fighting themselves (same logic as trading-plan's conflict flagging). Stock-bond correlation flips sign with the inflation regime (macro-regime), so "stocks + bonds = diversified" is regime-dependent — check which regime is live before crediting the hedge.
4. Options-adjusted truth. Sum short-put assignment obligations in dollars (that's shadow equity exposure at exactly the prices where it hurts); covered calls = sold right tail (the book's best names are the ones capped — say it); count net short-vol exposure across positions that all sell premium (one vol spike hits them simultaneously — that's a cluster too). Beta-weight option deltas into the book beta.
5. Stress scenarios (dollars, not percentages). Three concrete scenarios keyed to the book's actual tilts — e.g. rates +100bps, index −15% growth-led, the largest cluster's specific thesis breaking — and the estimated dollar drawdown of each, using stress correlations (→1 within risk assets; Longin & Solnik). Then the behavioral check: is the worst number one the user can sit through without panic-liquidating at the low? If not, the book fails the 11pm test at portfolio scale and the fix is pre-emptive trimming, not resolve (trading-plan field-card logic).
6. Cash and dry powder. Cash % vs the current market-health gate: zero cash in REDUCE_ONLY regimes means every new idea forces a sale (decision coupling); large cash in NEW_ENTRY_ALLOWED regimes is a position too — an implicit market-timing bet that should be owned as a decision.