Forensic Accounting — Are the Reported Earnings Real?

⚠️ Data Verification — Do This Before Any Analysis

Before running any analysis, always retrieve the latest market data for the ticker:

1. Fetch current price — use web search or ask the user for the live price, 52-week range, and market cap. Never assume a price from training data. (Market cap feeds Altman’s X4; everything else in this skill comes from the filings.)
2. Pull the filings, not a data feed — this skill needs at least two, preferably three to five, consecutive fiscal years of as-filed statements, plus their notes:
   • 10-K Item 8: income statement, balance sheet, cash-flow statement (for every index below you need year t and year t−1).
   • Revenue note (ASC 606): disaggregation, contract assets / unbilled receivables, deferred revenue, remaining performance obligations, any bill-and-hold or principal-vs-agent judgment.
   • Receivables / allowance note, inventory note (reserves, LIFO/FIFO changes), PP&E and intangibles notes (capitalised software, useful-life changes), capitalised contract costs (ASC 340-40), related-party note, stock-compensation note, commitments and contingencies.
   • Auditor’s report (opinion, critical audit matters, going-concern paragraph, tenure), Item 9A (internal control over financial reporting — material weaknesses), and 8-K Item 4.01 (auditor change) / Item 4.02 (non-reliance — restatement) for the last five years.
   • Earnings releases for the non-GAAP reconciliation (8-K Exhibit 99.1), and Form 4s for insider sales around results. Use as-filed figures; when a later filing restated a prior year, note both and compute the indices on the restated basis (the restatement itself is a finding).
3. State your data source — fill in the Data & Sources header (next section) so the origin, as-of date, retrieval path, and confidence of every figure are explicit at the top of the output.
4. Flag stale data explicitly — if live data is unavailable, display this warning before proceeding:

⚠️ Live data unavailable. The following analysis uses training-data estimates which may be significantly out of date. Verify all prices and metrics before making any decisions.

Never silently substitute training-data estimates for current prices. When in doubt, ask the user to paste the latest quote.

📋 Data & Sources Header — Open Every Output With It

The first thing in the output is this provenance block, filled in — never left as placeholders. It is the standard documented on the Data & Accuracy page and the first thing result-validator looks for:

Data & Sources
  As of:      <date the figures represent, e.g. 2026-06-30>
  Source:     <primary docs — SEC EDGAR 10-K/10-Q, company IR, FRED, exchange data …>
  Retrieval:  <pasted by user | web/tool retrieval | model memory>
  Confidence: <HIGH | MEDIUM | LOW>

• Retrieval: model memory must be paired with Confidence: LOW — memory is a placeholder until confirmed against a primary source.
• Mixed sources: list each with its own as-of date rather than blending them.
• Sources disagree on the same figure (a closing price, a deal term): prefer the primary source (SEC filing, company release, exchange data) over an aggregator, then the most recent dated one. Show both values with their sources and dates, say which you used and why, and never average them.
• Data the user pasted is reported as pasted by user; do not upgrade its confidence beyond what the user’s own source supports.

🌐 Output Language — --lang zh-TW

English is the default. With --lang zh-TW — or when the user writes in Traditional Chinese and asks for a Chinese answer — write the whole analysis in Traditional Chinese (繁體中文, Taiwan usage), keeping it machine-checkable:

• Translate every heading, narrative paragraph, table header and bullet. Give each metric’s English name in brackets the first time it appears — 自由現金流（Free Cash Flow, FCF） — so the reader can match it to the glossary and to English analyses.
• Keep in English tickers, figures and units, SEC form and item names (10-K Item 7, Form 4), the four Data & Sources field labels (As of, Source, Retrieval, Confidence — the Retrieval value may be written in Chinese, e.g. 使用者提供), and the Investment Signal box and JSON footer exactly as specified — field names and the BULLISH / NEUTRAL / BEARISH, HIGH / MEDIUM / LOW, BUY / HOLD / SELL, STRONG / MODERATE / WEAK values. Tools parse them.
• Title the invalidation section 論點失效條件（Thesis Invalidation）, and close with the disclaimer in Chinese: 免責聲明： 僅供教育用途，非投資建議。
• Use Taiwan financial terms: 殖利率、本益比、營業利益率、自由現金流、聯準會、選擇權、買權／賣權、除息日、預扣稅.

Overview

Every valuation in this catalog takes reported earnings as an input. This skill asks the question that comes first: how much of that earnings number is cash, how much is judgment, and do the patterns in the statements look like the patterns that preceded past manipulations and failures? It runs the academic screens that have a track record (Beneish’s M-Score, Altman’s Z, Sloan’s accruals anomaly), then the practitioner checks that catch what the models miss — revenue-recognition stretch, cost capitalisation, the widening gap between “adjusted” and GAAP earnings, stock compensation treated as free, and the governance tells (auditor changes, material weaknesses, restatements, related parties).

This is a screen for risk, not an accusation. A high reading means “the statements have the shape of trouble and deserve a skeptical read,” not “this company commits fraud.” Most companies that trip one flag have an innocent explanation (an acquisition distorts DSRI; a fast grower has high SGI by definition). The skill’s job is to compute the numbers exactly, explain each flag, and say which explanation the filings support.

Where it sits. financial-report-analyst reads the filing and gives an Accounting Quality Score; 10k-digest summarises it; bear-case argues the short thesis and mentions Beneish without computing it. This skill does the arithmetic and returns a Manipulation Risk Score (0–10, higher = more risk). Feed a high score into bear-case as evidence, check any surprising figure with fact-check, and read insider selling around the flagged periods with insider-trading.

1. Inputs

Quarterly mode. If the user asks about a recent quarter, compute DSO/DIO/DPO, unbilled receivables and the non-GAAP gap on trailing-twelve-month figures and the last eight quarters; run Beneish and Altman on fiscal years only (the coefficients were estimated on annual data).

2. Framework

Phase 1 — Beneish M-Score (8-variable model)

Compute each index from year t (latest) and year t−1. Show every input in a table so the reader can check the arithmetic.

Reference means are from Beneish (1999), “The Detection of Earnings Manipulation,” Financial Analysts Journal.

The model:

M = −4.84 + 0.920·DSRI + 0.528·GMI + 0.404·AQI + 0.892·SGI
          + 0.115·DEPI − 0.172·SGAI + 4.679·TATA − 0.327·LVGI

Report the M-Score against both cutoffs and name the one or two indices that drive it (largest coefficient × deviation from the non-manipulator mean). A score above the threshold driven only by SGI (a genuine fast grower) is weaker evidence than one driven by DSRI + TATA + AQI.

Computation rules.

• Securities in AQI = long-term investments (not cash). If the balance sheet shows none, use 0 and say so.
• Depreciation excludes amortisation of acquired intangibles where the filing separates them; Net PP&E excludes operating-lease right-of-use assets (ASC 842 inflates AQI otherwise — note it if the adoption year is in the window).
• A major acquisition in year t distorts DSRI, AQI and SGI. Flag it and, if the target’s pro-forma figures are disclosed, recompute on a pro-forma basis.
• The model was estimated on non-financial firms; do not run it on banks, insurers or REITs (Phase 7).

Phase 2 — Altman Z (distress, not fraud — but distress is a motive)

Choose the version by the company’s profile, state which you used and why:

Where X1 = working capital / total assets; X2 = retained earnings / total assets; X3 = EBIT / total assets; X4 = market value of equity / total liabilities; X4′ = book equity / total liabilities; X5 = sales / total assets. Z’’ drops X5 because asset turnover varies too much across non-manufacturing industries. For emerging-market issuers, Altman’s EM score adds a constant of 3.25 to Z’’ (zones > 5.85 / 4.35–5.85 / < 4.35).

Caveats to print. Large buybacks drive retained earnings (X2) negative for healthy companies — say so rather than calling them distressed. Asset-light software firms with big deferred revenue look worse on X1 than they are. Never apply Z to banks, insurers, or REITs.

Phase 3 — Accruals and cash conversion

Sloan accruals ratio — balance-sheet version (Sloan 1996):

Accruals_BS = (ΔCurrent assets − ΔCash) − (ΔCurrent liabilities − ΔShort-term debt − ΔTaxes payable) − Depreciation & amortisation
Accruals ratio_BS = Accruals_BS / Average total assets

Cash-flow version (cleaner; unaffected by acquisitions and reclassifications):

Accruals_CF = Net income − (Cash from operations + Cash from investing)
Accruals ratio_CF = Accruals_CF / Average net operating assets
   (NOA = (Total assets − Cash & investments) − (Total liabilities − Total debt))

Cash conversion. OCF / Net income for each of the last 3–5 years, and the cumulative ratio (Σ OCF / Σ NI) across the window. Cumulative ≥ 1.0 is healthy for a mature company; below 0.8 cumulatively is a flag. One bad year is noise; a declining trend while net income rises is the classic divergence. Also check where OCF comes from: a one-off boost from stretching payables, selling receivables (factoring), or supplier-finance programmes is not durable — look for it in the working-capital lines and the supplier-finance disclosure (ASU 2022-04).

SBC-adjusted free cash flow.

FCF          = OCF − Capex (including capitalised software)
SBC-adj. FCF = FCF − Stock-based compensation
SBC % revenue = SBC / Revenue

Report net dilution (diluted share count change, after buybacks) alongside it: buybacks that only offset SBC are a compensation expense paid in cash.

Phase 4 — Revenue recognition and working capital

Working-capital days, 3–5 years, vs. 2–3 peers:

DSO = Accounts receivable / Revenue × 365
DIO = Inventory / COGS × 365
DPO = Accounts payable / COGS × 365
CCC = DSO + DIO − DPO

Flag any of these rising more than 15% year on year without a disclosed reason (mix shift, new large customer with longer terms, acquisition). DSO rising while revenue accelerates is the core channel-stuffing pattern; DIO rising with flat sales signals obsolescence the reserve may not cover; DPO rising sharply flatters OCF temporarily.

Revenue-recognition red flags — check each, mark ✅ none found / ⚠️ present / ❓ not disclosed:

Phase 5 — Capitalised costs and the GAAP vs. non-GAAP gap

Capitalisation. For each category, report the amount capitalised in the year, as % of the related expense, and the trend:

Capitalising costs moves them from the income statement and from OCF into investing cash flow — so check capex including capitalised software, not just PP&E capex.

Non-GAAP vs. GAAP gap. For each of the last 3–5 years (and 8 quarters):

Gap % = (Non-GAAP net income − GAAP net income) / |GAAP net income|

Itemise the adjustments. Flag: a gap widening for three consecutive periods; “one-time” restructuring charges that recur in most years; SBC excluded and buybacks presented as returning capital; acquisition-related amortisation excluded while acquisitions are the growth strategy; adjustments the company added this year.

Phase 6 — Governance, audit, and related parties

Phase 7 — Sector exceptions (financials, insurers, REITs)

The standard models assume an operating company. For these sectors, do not compute the listed models — say they do not apply and use the substitutes:

Governance checks (Phase 6) apply to every sector. For these companies the score is built from Phase 6 plus the substitutes; renormalise the weights (Phase 8) and say so.

Phase 8 — Manipulation Risk Score (0–10, higher = more risk)

Score each component 0 (clean) to 10 (severe) on the anchors, then weight:

Manipulation Risk Score (MRS) = Σ weight × component score
Box Score                     = 10 − MRS

Hard floors on MRS (state any that fired):

• Going-concern paragraph, auditor resignation with a disagreement, or a Big-R restatement in the last 3 years → MRS ≥ 7.0
• Unremediated material weakness, late 10-K filing, or an open SEC enforcement inquiry → MRS ≥ 6.0

Missing data. If a component cannot be computed (one year of data only, sector exception), drop it and renormalise the remaining weights to 100%, list what was dropped, and cap Confidence at MEDIUM. Never score an uncomputed component as 0.

3. Output Format

1. Data & Sources header (above), listing the filings (form, fiscal year, accession number) used
2. Executive Summary — three sentences: the MRS and what drives it, the single most important flag, what to do (monitor / dig further / avoid until explained)
3. Beneish table — eight indices with inputs, t and t−1 values, the M-Score vs. −1.78 and −2.22, and the driving indices
4. Altman table — version used and why, X1–X5, Z, zone, and any caveat (buybacks, deferred revenue)
5. Accruals & cash — Sloan ratio (both versions), OCF/NI by year and cumulative, FCF vs. SBC-adjusted FCF, SBC % revenue, net dilution
6. Revenue & working capital — DSO/DIO/DPO/CCC trend vs. peers; red-flag checklist (✅ / ⚠️ / ❓)
7. Capitalisation & non-GAAP — capitalised amounts and trend; GAAP vs. non-GAAP bridge by year
8. Governance & audit — auditor, CAMs, going concern, MW, restatements, related parties, insider sales
9. Manipulation Risk Score — component table, hard floors fired, MRS, box Score = 10 − MRS
10. Thesis Invalidation — what would reverse the call (below)
11. Investment Signal Block — the standard box (below)

Example

User: Run a forensic accounting check on ZEPH (fictional ticker, FY2024–FY2025 10-Ks pasted)

The assistant fills the Data & Sources header (two 10-Ks, pasted by user, HIGH),
then computes the Beneish indices. Hypothetical inputs: revenue grew from $800M to
$1,000M (SGI 1.25) while receivables grew from $120M to $200M, so DSRI =
(200/1,000) / (120/800) = 1.33 — well above the non-manipulator mean of 1.03.
Net income of $90M against OCF of $40M on $1,600M of assets gives TATA = 0.031,
at the manipulator mean. With GMI 1.02, AQI 1.10, DEPI 1.05, SGAI 0.95 and
LVGI 1.00, M = −4.84 + 1.227 + 0.539 + 0.444 + 1.115 + 0.121 − 0.163 + 0.146
− 0.327 = −1.74: above both cutoffs, driven by DSRI and SGI.

Z'' (ZEPH is a software company) sits in the safe zone, so there is no distress
motive. The cash-flow accruals ratio is +14% and cumulative OCF/NI over three years
is 0.62. The revenue note shows unbilled receivables up 140% against revenue up
25%, and capitalised commissions growing faster than revenue. The auditor is
unchanged and there is no material weakness.

Components: Beneish 8, accruals 8, revenue 8, capitalisation 6, governance 2,
distress 1 → MRS 6.1, box Score 3.9, BEARISH, MEDIUM confidence (two years only;
peer DSO not supplied). Next steps: `fact-check` the receivables figures,
`earnings-call-analysis` for management's explanation of unbilled growth,
`bear-case` to build the full short thesis.

Notes

• Screens, not verdicts. Beneish’s model catches a meaningful share of later-revealed manipulators but also flags many innocent firms; Altman’s Z predicts distress, not fraud. Use the language “consistent with” and “warrants scrutiny,” never “the company is manipulating earnings.”
• Never invent figures. Every index needs two years of exact line items; if they were not retrieved or pasted, mark the index “not computed” rather than estimating it from memory.
• Show the arithmetic. Print the inputs to every ratio. A forensic report the reader cannot re-check is worthless — and fact-check will re-check it.
• Context beats thresholds. An acquisition, a change in fiscal year, ASC 606 / ASC 842 adoption, or a business-model shift (hardware to subscription) breaks year-over-year indices. Identify the event and, where possible, compute on a like-for-like basis.
• Fast growers trip SGI and DSRI honestly. Compare DSO with peers and with the company’s own history before calling it stuffing.
• Short-seller reports are allegations; test their specific claims against the filings and report which are supported, contradicted, or unverifiable.
• Run before: 10k-digest or financial-report-analyst (to orient in the filing). Run after: bear-case (to argue the risk), fact-check (to verify any figure that drives the score), insider-trading (for sales around flagged periods), credit-check (if the Altman reading is grey or distressed).

Thesis Invalidation

After delivering the analysis signal, specify what would reverse it:

If signal is BULLISH (low manipulation risk) — thesis breaks if:

• DSO or unbilled receivables rise more than 15% year on year without a disclosed reason in the next 10-Q, or cumulative OCF/NI falls below 0.8
• The auditor resigns, a material weakness is disclosed, an 8-K Item 4.02 non-reliance notice is filed, or a 10-K is filed late (NT 10-K)
• The non-GAAP gap widens for a third consecutive period or a new capitalisation policy appears
• A credible short-seller report or SEC inquiry makes specific claims the filings cannot rebut

If signal is BEARISH (high manipulation risk) — thesis breaks if:

• The flagged receivables or contract assets convert to cash in the next one to two quarters (OCF catches up with net income)
• The driver of the M-Score is explained by an identified event (acquisition, ASC 606 transition) and the like-for-like recomputation falls below −2.22
• A material weakness is remediated with a clean ICFR opinion, or a new auditor issues a clean opinion without restating
• The company narrows its non-GAAP adjustments or begins expensing previously capitalised costs

Re-run this analysis when:

• [ ] Next 10-K is filed (the models are annual) — and each 10-Q for the working-capital and non-GAAP checks
• [ ] Any 8-K Item 4.01 (auditor change), 4.02 (non-reliance), or 5.02 (CFO departure)
• [ ] An NT 10-K / NT 10-Q late-filing notice, or a short-seller report is published
• [ ] Price moves ±15% from current level
• [ ] Material news event (acquisition, leadership change, regulatory decision)

Standard Signal Output

This skill measures reporting risk, not valuation or direction, so state the mapping: Signal: BULLISH means the reported earnings look trustworthy enough to use as an input for valuation; NEUTRAL means specific items need monitoring before relying on them; BEARISH means treat the reported earnings as unreliable until the flags are explained. Action: SELL here means “do not underwrite a position on these numbers” — read whether the stock is cheap from stock-valuation, not from this block.

All analysis concludes with this standardized block:

In the output, the box below is followed immediately by this machine-readable footer — the same values as the box, filled in (no placeholders, no comments), so tools can read the result without parsing the box:

{"investskill": "1", "skill": "forensic-accounting", "ticker": "<TICKER, or null for a portfolio / market-wide view>", "as_of": "<YYYY-MM-DD from Data & Sources>", "signal": "<BULLISH | NEUTRAL | BEARISH>", "confidence": "<HIGH | MEDIUM | LOW>", "horizon": "<SHORT-TERM | MEDIUM-TERM | LONG-TERM>", "score": 0.0, "action": "<BUY | HOLD | SELL>", "conviction": "<STRONG | MODERATE | WEAK>", "retrieval": "<pasted by user | web/tool retrieval | model memory>", "sources": ["<each primary source, dated>"]}

╔══════════════════════════════════════════════╗
║              INVESTMENT SIGNAL               ║
╠══════════════════════════════════════════════╣
║ Signal:      BULLISH / NEUTRAL / BEARISH     ║
║ Confidence:  HIGH / MEDIUM / LOW             ║
║ Horizon:     SHORT / MEDIUM / LONG-TERM      ║
║ Score:       X.X / 10                        ║
╠══════════════════════════════════════════════╣
║ Action:      BUY / HOLD / SELL               ║
║ Conviction:  STRONG / MODERATE / WEAK        ║
╚══════════════════════════════════════════════╝

Score Guide: 8.0–10.0 Strongly Bullish | 6.0–7.9 Moderately Bullish | 4.0–5.9 Neutral | 2.0–3.9 Moderately Bearish | 0.0–1.9 Strongly Bearish Confidence: HIGH (strong data, clear signals) | MEDIUM (mixed signals) | LOW (limited data, conflicting signals) Horizon: SHORT-TERM (1 week–3 months) | MEDIUM-TERM (3 months–1 year) | LONG-TERM (1+ years)

Note: The box Score is 10 − Manipulation Risk Score, so higher manipulation risk produces a lower Score and a more bearish signal (MRS 7.0 → Score 3.0 → BEARISH); always print the MRS itself next to the box. Confidence is at most MEDIUM when fewer than three fiscal years were analysed or any component was dropped, and LOW when the figures came from model memory.

Disclaimer: Educational analysis only. Not financial advice.