heurist-finance: Skill for Claude Code

Fundamental Forensics

Shared plumbing, integrity rules, and output contract: ../_shared/analyst-core.md. This skill answers one question: are the reported earnings real, and is the market about to find out otherwise?

Why this beats reading the income statement

Reported EPS is the single most managed number in markets. The documented anomalies all live in the gap between earnings and cash: Sloan (1996) found top-decile-accrual firms underperform bottom-decile by ~10% over the next year; Dechow's AAER work shows restatement firms cluster on the same handful of precursors years in advance. A smart agent's edge here is that the inputs are free (XBRL) and most market participants still don't look.

The workflow

1. Pull 8–12 quarters of the core facts via sec xbrl_fact_trends (exact XBRL names; dedupe restatements by period-end): NetIncomeLoss, NetCashProvidedByUsedInOperatingActivities, AccountsReceivableNetCurrent, Revenues (or RevenueFromContractWithCustomerExcludingAssessedTax), InventoryNet, Assets, PaymentsToAcquirePropertyPlantAndEquipment. Cross-check headline figures against yahoo equity_overview.
2. Compute the diagnostics (in Python from the pulled JSON, not by eyeball):
   • Cash conversion: CFO ÷ NI, trailing 4q. < 0.8 sustained over ~3 years is the classic pre-restatement cluster (Dechow et al. 2011) — especially paired with rising DSO.
   • Total accruals / assets = (NI − CFO) ÷ avg assets. Sustained > ~5% is the Sloan red zone. Split working-capital accruals from long-term/investment accruals (capex, acquisitions) — Richardson et al. (2005): the long-term kind predicts underperformance harder.
   • DSO trend = AR ÷ (revenue/91). Rising DSO while revenue accelerates = revenue being pulled forward (channel stuffing signature). Same logic for inventory days on product companies.
   • Growth-vs-cash test: revenue growth > 30% with CFO/NI < 0.5 for 2+ years is the canonical blow-up profile (this exact pattern preceded most famous frauds).
   • NI-vs-OpInc bridge: if net income grows faster than operating income, pull the non-operating lines (InvestmentIncomeInterest, OtherNonoperatingIncomeExpense) and attribute the gap — interest on a cash pile is real income but not operating leverage, and a thesis priced off the NI line inherits the rate cycle.
3. Read the qualitative tripwires via sec filing_timeline + filing_diff on the last two 10-Ks: auditor change, new related-party disclosures (> ~5% of revenue is serious), CFO departure, changes in revenue-recognition language, growing "unbilled receivables." filing_diff on the risk-factors section is the cheap way to catch what management's lawyers just added.
4. Check who's acting on it: sec insider_activity (are officers selling into the growth story?) and short interest context (yahoo equity_screen screen_name=most_shorted_stocks, equity_overview).
5. For cheap stocks specifically, run the Piotroski F-Score logic (9 binary points: profitability, leverage/liquidity direction, efficiency) — but only after confirming cheapness (bottom-of-range P/B or EV/EBITDA vs the name's own history and sector); Piotroski's spread lives inside the cheap bucket, and F-scoring a richly valued name proves nothing. Piotroski (2000): high-F cheap stocks beat low-F cheap stocks by ~7.5%/yr in-sample. The point isn't the score — it's that cheap + deteriorating = trap, cheap + improving = candidate. Never let "it's cheap on P/E" stand alone.