How Filing.fyi reads a filing.

1. What we pull, and from where

• SEC EDGAR. Annual reports (10-K), quarterly reports (10-Q), current reports (8-K), insider transactions (Form 4), and institutional holdings (Form 13F). We pull directly from the SEC. We do not buy aggregated filing data from a vendor.
• yfinance. Delayed (15+ minute) closing prices, basic ratios, market capitalization, and historical price series, plus the short-interest figures and closing option prices behind the market signals (section 5). Every yfinance-derived field is free for all users; we do not paywall it and we do not market on it.
• FRED (St. Louis Fed). Macro indicators rendered in the daily Market Pulse band — VIX, 10-year Treasury yield, Shiller CAPE proxy, equity risk premium.
• News headlines. RSS-licensed headlines only. We do not scrape news bodies and we do not store article text.

2. Forensic scores — what they are and where they came from

The forensic suite on every ticker page is a set of mechanical, peer-reviewed scores. They are computed from the company's own filed numbers. They are not predictions.

These methods were developed for academic research, sometimes decades ago. They can produce inconsistent results across industries, can be miscalibrated against today's accounting standards, and can be wrong. Treat them as one input among many, not as a proof of anything.

3. Verdict labels — what they actually mean

Filing.fyi uses four verdict labels, plus a directional chip. The labels are editorial categorizations, not regulated ratings. They describe what the filings show about a company's financial health and accounting quality; they are not signals to buy, hold or sell, and Deep Value does not mean "buy" any more than Red Flags means "sell".

• Deep Value — the forensic scores are strong across the board (health score 75 or higher). It describes financial health and accounting quality in the filings, not a comparison of price with value. Does not mean the security is a bargain or that it will appreciate.
• Fairly Valued — the scores are solid without standing out (health score 55–74). Does not mean the price is correct or that the company is safe.
• Watch — the scores are mixed (health score 40–54). Does not mean you should take any action.
• Red Flags — the scores are weak (health score below 40). Does not mean fraud has occurred or that the company will fail.

The health score turns each available score (Beneish, Altman, Piotroski, Fog) into points and averages them on a 0–100 scale; a deeply negative Altman reading caps it at 35. When a company has no computed scores, the label comes from the AI analysis of its latest filing; when there is neither, the page says "Not yet scored" and shows no label. How to read a company page walks through it with an example.

We deliberately do not use "Buy," "Hold," or "Sell." Those words imply a personalized recommendation, which Filing.fyi cannot make and is not licensed to make. See our Terms, §3–§5.

4. The synthesis — how the prose gets written

The "Synthesis" section on each ticker hub is a multi-paragraph readthrough of the most recent material filing. It is generated by a third-party large-language-model API, with provider fallbacks, using a fixed editorial prompt grounded in the filing text and the computed forensic scores. The prompt is calibrated on the framework of Howard Marks, Benjamin Graham, and Ray Dalio — not on social-media sentiment.

No human edits the synthesis prose. The model can be wrong, can be internally inconsistent, and can hallucinate. Treat any passage it attributes to a filing as paraphrase, not quotation, unless you check it against the document on EDGAR. Each narrative's byline names the filings it was written from and the date it was written.

5. Market signals — short interest and options

Below the synthesis, each company page shows two readings of how traders are positioned in the shares. They come from trading data, not from the filings, and they are not part of the verdict or the health score.

• Short interest. Shares sold short and not yet bought back, as reported to FINRA for two settlement dates a month and published about a week and a half later (via yfinance). We show the share of the float sold short (or of all shares outstanding when no float figure is available), the change since the previous report, days to cover (shares short divided by average daily volume), and every report we have stored. Bands: under 5% low, 5–10% moderate, 10–20% high, 20% and up very high.
• Options. Computed by Filing.fyi each weekday evening from that day's closing option prices (bid/ask midpoint, else last trade, via yfinance). Implied volatility uses the Black-Scholes model on the strikes nearest the share price for the expiration closest to 30 days out (7–60 days), averaging the call and the put and ignoring dividends and early exercise. The typical move priced in is implied volatility × √(days to expiration ÷ 365): about a two-in-three range under the model's assumptions, with no direction. Actual volatility is the annualized standard deviation of the last 21 daily log returns. The put/call ratio sums open interest over up to four expirations 7–60 days out. Skew compares the implied volatility of a put and a call each about one typical move away from the share price. Once 20 trading days are on record we also say where today's implied volatility sits in the company's own history.

The bands and wording on these readings are plain-English reading aids, not research-calibrated cutoffs. We store derived numbers only; we never publish an option chain. See the lessons on short interest and the options market.

6. Refresh cadence

• Forensic scores + synthesis. Recomputed when a new 10-K, 10-Q, or material 8-K is detected on EDGAR for a covered ticker. Cached for at least 7 days.
• Market Pulse. Recomputed daily at 5:00 PM ET on weekdays.
• Delayed price data (yfinance). Recomputed daily at 6:00 PM ET on weekdays. Free for all users.
• Top losers screen. Recomputed daily at 5:30 PM ET on weekdays.
• Short interest and options readings. Recorded daily at about 6:20 PM ET on weekdays from closing data; short interest itself changes twice a month.
• News headlines. Pulled on page render with a short cache.

We do not push intra-day refreshes and we do not surface live ticks. If you need real-time pricing for execution, use your broker.

7. What we do not do

• We do not provide personalized investment advice. The custom-query feature has hard-coded refusals for "should I buy / given my $X / for my IRA" patterns.
• We do not market on price moves. You will never see "rallied," "surged," or "up X%" in our list pages, hero copy, or pricing page.
• We do not bulk-export yfinance-derived data, and we do not publish raw option chains.
• We do not trigger watchlist alerts on raw price moves, short interest, or options readings — only on filing events and forensic-threshold crossings.
• We do not edit the substance of user comments. We approve, reject, or shadow.
• We do not pretend to be a registered investment adviser. We are not.

8. Limitations and known risks

• The forensic scores were developed on different decades' data and may be mis-calibrated for current accounting standards (IFRS / new SFAS, restated comparables).
• The synthesis prose is LLM-generated and can be wrong. Always verify a load-bearing claim against the underlying filing.
• The Buffett Indicator and Shiller CAPE rendered in the Market Pulse band are approximations. The true CAPE requires a 10-year inflation-adjusted earnings history we do not currently maintain.
• Short interest is published twice a month with a delay of about a week and a half, and it does not say who is short or why; some of it is hedging. Option-implied numbers rest on a model whose assumptions markets break, and they can swing on a few trades when few contracts are open.
• Insider transaction data from Form 4 can be filed late or amended; we do not retro- actively correct.
• We do not cover every public US ticker. Coverage is rolled out per-ticker and gated on the uniqueness and quality of the generated synthesis.

9. Corrections

If you believe a score, a citation, or a sentence in a synthesis is wrong, email hello@filing.fyi with the ticker, the URL, and what you think the correct answer is. We will look at it. If the underlying inputs from the filing are wrong, we will rerun the pipeline; if the LLM produced a hallucination, we will note the correction and adjust the prompt.

Methodology changes are reflected by updating the "Last reviewed" date above and, when material, by a note in the RSS feed.