How FinPile Scores Stocks — Our Methodology

Every score, rating, and flag on FinPile is computed by fixed, published rules — deterministic code, not AI judgment. AI writes the plain-English explanation after the scoring is done, and it cannot change a single number.

Most AI stock tools work the other way around: they hand your ticker to a language model and let it improvise an opinion. We think that's the wrong architecture for financial analysis, because you can't audit an improvisation. You can audit a rulebook. This page is ours.

That includes the one word at the top of every stock page. FinPile's Buy / Hold / Sell rating is a fixed function of a five-point checklist and the measured odds for the stock's class — the definition is below, and every call is scored on the Ledger.

Every rule in our engines must be statable without naming a company, a country as a prejudice, or a sector as a prejudice. No ticker is special-cased. No market gets a thumb on the scale.

That doesn't mean every stock is treated identically — a micro-cap with three reverse splits should read differently than a mega-cap with a buyback program. It means the differences flow from stated, universal conditions any stock could trigger: "listings quoted in pence get currency normalization," "sectors on this named list get asset-heavy valuation treatment," "any company whose share count fell more than 50% in a year gets that trend excluded as a probable restructuring artifact." Same trigger, same treatment, every time.

Financial Health — 11 signals, one rating (12 for financial institutions). Profitability, cash-flow quality (is net income backed by operating cash?), free-cash-flow status, returns, growth, liquidity, leverage, interest coverage, share-count trend, and equity status. Banks and insurers are scored on bank-appropriate metrics — industrial solvency ratios are excluded as not meaningful, and ROE joins the set. Each signal passes, fails, sits in a neutral band, or is marked not meaningful with its reason. The weighted total maps to five ratings, Strong to Distressed — and the badge shows the score, the fail count, and an edge qualifier when a rating sits at its band's boundary, because a Strong at 4.0 with two fails and a Strong at 11.5 with none are different animals and should look it. A perfect sweep — every signal passing, no caveats — earns a "flawless" mark; it's rare, and it describes the snapshot, not the future. A company whose revenue is collapsing gets its rating capped no matter how pretty the other signals look.

Valuation — multiples with a loss-maker rule, a band line, and a refusal. Six multiples judged against their sector medians, with the stock's own 5-year range shown beside each (rule v2, Oct 2026: a multiple reads elevated when it exceeds 1.2× its sector median or a fixed sanity ceiling — P/E 40, P/S 10, EV/EBITDA 25, P/B 6 — and cheap when it sits under 0.8× the median or a sanity floor; without a sector median the fixed absolute thresholds apply. Under the earlier rule a multiple had to clear both the absolute threshold and the sector multiple, which left most stocks in the Fair range by construction; the next weekly scan re-scores every stock under v2, and the Ledger lists each valuation word that moves). Earnings-based measures lead when real earnings exist; loss-makers are judged on revenue-based multiples and labeled as such. Fewer than two usable multiples → "insufficient data," not a guess. The valuation band shows the price zones where our own verdict would flip — and when the current price sits far outside every zone (beyond ~2× the boundary), we refuse to draw the band and say why: at that distance, the market is pricing expectations far beyond reported results, and dollar zones would imply a precision the arithmetic doesn't have.

DCF Sensitivity — assumptions shown, spread as the answer. We display a discounted-cash-flow grid across terminal-growth and discount-rate assumptions, with every input disclosed (WACC, beta, growth path — from our data provider's model) and our own recomputation reconciled against it within 2% before anything renders. The spread across the grid is the finding; we never label any cell "fair value." The module gates itself off for financial institutions, negative-cash-flow companies, and stale computations — a negative "intrinsic value" is an artifact, not a valuation.

Implied Expectations — the DCF, run backwards. A forward DCF guesses the future to produce a value, which is why we refuse to print one. A reverse-DCF inverts the arithmetic instead: it takes today's price as given and solves for the annual free-cash-flow growth rate that price already implies, then shows it beside what the company has actually delivered — the gap between implied and realized is the finding. It rests on three fixed, deliberately disclosed assumptions, rendered with every result: a single transparent 10% discount rate (not a per-stock WACC estimate, which would add precision the inputs don't support), a 10-year horizon, and 2.5% terminal growth thereafter, grown from trailing-twelve-month free cash flow. Reverse-DCF is more honest than forward-DCF but it is not assumption-free, so you can always see exactly what the number rests on — a different discount rate implies a different growth rate. It gates for the same cases the forward DCF does: banks, insurers, REITs, pre-revenue companies, and negative trailing free cash flow, where a growth rate cannot be solved from a negative base at all. This is a separate lens from the valuation verdict, not an input to it: it feeds no rating, no cell, no scanner, and no score. It reveals the expectations already priced in; it never estimates what a company is worth, and whether the implied growth is plausible is your judgment to make.

One lens, two rules — stated side by side so the difference is deliberate, not drift. The implied growth rate is read by two modules that ask different questions, and each has its own rule. The evidence tally asks the valuation question — is the price assuming more than the company has shown? — so it reads the rate relatively: implied growth more than 5 points above every recent realized figure (last-year revenue, 3-year revenue CAGR, TTM free-cash-flow change) leans unfavorable; implied growth below every one of them leans favorable; anything between is neutral. The "Into earnings" panel asks the setup question — how demanding is the assumption in absolute terms? — so its badge tier reads the rate absolutely: implied free-cash-flow growth of 15%/yr or more is a High bar, 5–15% a Moderate bar, below 5% or negative a Low bar. A decade-long ≥15% assumption is demanding regardless of what recent quarters did, so the relative comparison becomes the modifier sentence there, honest in both directions: "a rate the company has recently been clearing" when delivery sits above the implied rate, "above what the company has recently delivered" when it sits below. A stock can therefore lean favorable in the tally and still carry a High bar into a report — both statements are true, about different things.

Dilution Risk — a 100-point score built from SEC filings. We read EDGAR directly: shelf registrations, offering documents, split-adjusted share-count history, reverse-split records, insider Form 4s. Dilution is measured split-adjusted, so a reverse split can't launder a rising share count. Debt offerings are classified and excluded — borrowing is not printing shares. Insider transactions are classified by their filing codes: open-market purchases are distinguished from option exercises, scheduled-plan sales, and awards, and transactions pre-dating a company's listing are flagged rather than counted.

Technical read — levels, structure, and history. A price ladder of the monthly and weekly levels the market has actually respected, with break/hold/reclaim detection; a structure read from trend alignment and level interaction; seasonal history over ten years with win rates, ranges, and flags on the months where an earnings date — not the calendar — drives the average; and gap-behavior statistics (how this stock's past gaps resolved, by direction and by earnings context, every rate shown as "X of N"). Pattern detection names what the combined picture resembles — serial diluter, low-float mover, value trap, and others — each with published trigger conditions.

Election History — calendar windows, counted honestly. Around each US election in a stock's available history we measure the ±60-trading-day window — total return, drawup/drawdown, realized range — each beside SPY over identical dates, and the volatility comparison is testable rather than asserted: election-window range is shown against the stock's own average across all ordinary 120-day windows. Every figure carries its n; medians with ranges, never bare averages; fewer than three elections gates the view with the reason. The four-year-cycle table averages only the market's record; a stock below five complete cycles gets its years listed, not averaged. Administrations render only as an undigested chronological list with a named macro confound per window — no "under party X" average exists at any n, because five experiments is a story, not a statistic. The module is context only: it feeds no verdict, no screener, no scanner.

Catalysts — type and materiality, never direction. We classify what happened and how big it is. We retired "positive/negative catalyst" labels deliberately: whether news is good for shareholders is the market's call, and press releases are written to sound like good news. Where an event's mechanics are checkable — a dilutive raise on a thin float, a dividend cut — we name the mechanism; we never grade the story.

The Scanner — cells, not picks. Weekly, we place every universe stock into a grid cell by health × valuation — descriptive names ("Quality at a fair price," "Value-trap territory"), never rankings. Trap cells render as prominently as attractive ones. After every run, an assertion verifies each stock's cell matches its own displayed chips — the cells and the pages can never disagree silently. Empty cells render with their counts, because "no quality is visibly cheap this week" is a finding.

FINRA publishes short interest twice a month. The level (how much of the float is short) says how crowded the position is; the change across the last two settlements says whether it is being added to or unwound. Both are read, and they have strictly separate jobs: the level owns the direction of the lean, the change owns only how strongly it is stated. A change can never flip an unfavorable read favorable, or the reverse.

Change tiers, measured over one settlement cycle and share-basis verified: flat under 10% · modest 10–25% · steady 25–40% · rapid above 40%. If the share basis moved across the seam — a split, an exchange, a restatement — the two reports are not comparable and the change is reported unread; the lean is then the level's alone. Nothing is estimated across a basis break.

The change against price, over the same window

Price is measured on the same two settlement dates as the short change — the close on the newer settlement against the close on the older one, never "today versus then". Up is more than +5%, down is less than −5%, anything between is flat. The four readings are structure descriptions, not outcomes; a flat-price window gets no reading at all.

• •Building + price up → “shorts building into strength”
• •Building + price down → “shorts pressing a working thesis”
• •Covering + price up → “shorts covering into strength”
• •Covering + price down → “shorts stepping back as price falls”

The borrow extension

The FINRA print is already about nine business days old when it publishes, so the daily borrow book extends it: the fee today against the fee as it stood on the newer settlement date. More than 25% higher reads “borrow cost has risen since — the pressure looks ongoing”; more than 25% lower reads “has eased since”; anything between reads “steady since”. This covers Interactive Brokers' lending book only, which is stated every time, and when the book has no row that old the sentence is omitted rather than guessed.

Level × change → the lean

Generated from the engine's own table at render time (spec short-change-1.0) — bold cells are the ones the change moved. No cell in this table produces a favorable lean: covering makes a crowded short read less strongly, never well.

A rapid build or cover on a watched ticker also raises one Guardian notice per settlement — a structural event, carrying the reading, the share counts, both settlement dates and the borrow line, and never an outcome. The event class arms only for settlements published after it shipped: already-public prints are recorded silently rather than notified after the fact.

When a stock's short interest reaches 10% of its shares outstanding, the page shows what happened to past stocks with the same short structure — a count over stored episodes, never a statement about the stock you are reading. These are the rules behind it.

It starts in September 2021. The stored series begins with entries in Sep 2021, so the early-2021 short-interest episodes are not in the study and no rate on the site describes them. Every line states the start date beside its sample size.

Short interest is measured against shares outstanding, not float. No feed carries historical float, so a float-based percent could not be reconstructed for past episodes without inventing it. Percent of outstanding is smaller than the float percent you may see quoted elsewhere; the card shows its own arithmetic — shares short ÷ last filed share count, with that filing's date.

Entry is the day after FINRA published the reading. Short interest settles twice a month and is published about nine business days later. Measuring from the settlement session would use a number nobody could have known, so every episode is measured from the first session after publication.

There is a null model. Every rate is shown beside lightly shorted stocks — under 5% of shares outstanding, same universe, same windows. A rate without that comparison says nothing, so a line missing it is dropped rather than shown alone. Delisted listings are included on purpose: dropping the companies that died would flatter every number.

Days to cover runs the other way from the folk theory. In the stored episodes, stocks with under two days to cover rose 40%+ more often than stocks with ten or more — the opposite of the idea that a crowded, slow-to-exit short is the primed one. And the most crowded, slowest cell — short interest above 30% of outstanding with ten or more days to cover — is not where the rises concentrate; it carries the study's higher rate of 20% falls. We publish that because it contradicts the story the pattern is usually told with.

A rate that does not survive being cut in two is not shown as a rate. Each cell is re-counted over the first and second half of its episode sequence. A cell that moves 10 points or more between halves is tagged "unstable across periods", and a cell with fewer than 30 episodes reads "too few to read" with no number at all. The study feeds no score, no verdict, no ranking, and no watchlist — it is context, and it never describes what any stock is about to do.

Three additions sit alongside the verdicts without changing any of them. None feeds a score, a state, or an alert, and none of them is a composite: there is no summed number, no 0–100 rating and no letter grade anywhere in this layer.

Percentile rank — standing, dated to a snapshot. Ranks are computed once per snapshot cycle from the weekly universe scan, never per page load, and every rendering states the snapshot's own as-of date. A stock's health score is compared against every scanned company with a readable cell, universe-wide and within its sector, and the sector sample size n is always shown — a percentile over eleven peers is a different claim from one over four hundred. A ticker outside the snapshot says so ("no universe rank") rather than borrowing a neighbour's. Only quality factors are ranked: cheap-versus-the-market is not cheap-versus-its-own-history, so the valuation lean deliberately carries no percentile.

Position in band. A verdict word covers a range, so each factor chip also shows where inside its own band the reading sits — a health score at the bottom of a band and one at the top render distinguishably. The position is read from the same values the verdict was computed from (the health score against its band edges, the multiples against their own five-year ranges, short interest against the heavy and slow-to-unwind lines), so it cannot drift from the verdict it annotates.

Situation profiles. A factor pattern says more than any average: strong books at a high multiple is a different situation from middling everywhere, even though a composite would score them alike. So the pattern gets a name, assigned by the fixed rules below in code (not AI), evaluated top-down with the first match winning, one profile per stock. The names describe a situation and never prescribe an action; the whole table is scanned for directional wording at render time and withheld rather than reworded if it ever fails. A stock matching nothing cleanly reads "Mixed profile" honestly.

Health bands are the engine's own five ratings; the valuation lean is the valuation gauge's own verdict; "dilution/risk elevated" means the risk engine's High or Severe band; "short structure heavy" is the short module's own heavy-against-float line. Nothing in this table recomputes any of them.

Company fundamentals and prices come from Financial Modeling Prep. Dilution and insider analysis comes from SEC EDGAR — the primary source, read filing by filing. (During one investigation we found major free aggregators displaying share counts inflated ~25× because they'd missed a reverse split. That's why we go to the filings.)

This is where most tools quietly fail, so we're explicit:

• •A metric that isn't computable is marked N/M with the reason — never silently substituted with a look-alike.
• •"Not reported by provider" and "unavailable" mean different things here, and we use the right one.
• •Data artifacts are gated, not scored: a literal-zero interest coverage on a company with billions in operating profit is a feed artifact, not distress; a 70% share-count drop on a new listing is restructuring, not buybacks. Excluded, with the reason shown.
• •Newly listed companies get a visible low-coverage qualifier — "based on 5 of 11 usable signals" — and time-window features (seasonal history, 5-year ranges, insider tallies) gate themselves when the history doesn't exist rather than implying one.
• •Financials older than ~6 months trigger a staleness warning. Every analysis states its statement basis — "TTM through [quarter-end], filed [date]" — and when a company reports earnings, the analysis recomputes on the event rather than waiting out a cache timer.
• •Percentages are never shown over immaterial bases: a 4,000% position change on a rounding-error holding renders as the share counts instead, because the percentage would be noise wearing precision.

Every render runs internal consistency checks: signal counts must reconcile with the audit table, operating cash flow must agree in sign everywhere it appears, free cash flow must equal its own components, the P/E must equal price over the displayed EPS. A failed check degrades the affected figure to "not meaningful" with the failure stated — we would rather show a labeled gap than a number we can't stand behind. Implausible provider values are bounded out before scoring, so a corrupt figure never becomes a verdict input.

After the deterministic engines produce every number, an AI model writes the narrative explanation — under constraints: it receives only pre-computed results, its output is appended after code-written lead sentences, and a post-generation filter strips claims the data doesn't support. If the AI provider went down entirely, every score on FinPile would be identical; you'd just read tables instead of prose.

We maintain a pinned regression suite — real tickers spanning mega-caps, foreign listings, serial-diluter micro-caps, reverse-split cases, and sparse-data new listings — with frozen expected outputs, plus synthetic fixtures that run the same gauge code as production. Every engine change is verified against the full suite before it ships. The methodology is versioned; when a rule changes, the change and its reason are recorded, and this page's version number moves.

• •No price predictions or price targets. Nobody reliably predicts short-term prices; tools claiming to are selling confidence, not analysis.
• •No ratings by opinion. FinPile does publish Buy / Hold / Sell — but by the rule above, from filed facts and measured frequencies, with the definition public and every call scored. No analyst, no model and no one here can move a rating by hand.
• •No direction labels on news. Type and materiality only — the market grades the story.
• •No fair-value points. Bands, spreads, and refusals — never a single number pretending the future is priced.
• •No hidden AI opinions. If a number appears on FinPile, there is a rule behind it that this page describes.