SentiSense Methodology

SentiSense covers publicly listed US equities, with emphasis on S&P 500 constituents, and is actively expanding. We track companies, institutional investors (13F filers), corporate insiders (Section 16 reporters), and US political officeholders subject to STOCK Act disclosure, with derived signals at the sector and index level.

Out of scope today: private company valuations, cryptocurrencies, FX, commodities, and fixed income.

We ingest from public financial press, SEC EDGAR filings, government disclosure portals, company investor relations pages, and curated social sources (selected public communities and partner channels with consent). We license market data and fundamentals from commercial providers.

We deliberately do not ingest paywalled content we are not licensed to redistribute, leaked material non-public information, or anonymous tip channels.

SentiSense uses AI across the pipeline. Precisely:

AI handles sentiment classification, entity resolution, news clustering, structured extraction from unstructured text, and the drafting of summaries and research reports from structured inputs.

AI does not generate investment recommendations, set price targets, take directional positions on covered companies, fabricate data, or rewrite primary source quotations.

Our research is built to present multiple perspectives. When a company reports earnings, we cover the bull case and the bear case. When institutional flows shift, we describe what that implies and what it does not. The system is designed to help readers think, not to think for them. Each report links its factual claims to source documents. If we cannot cite it, we do not claim it.

Sentiment scores are real-valued in the range of minus one to plus one. Aggregated sentiment is a reliability-weighted mean of constituent document scores. Sources are tiered by reliability, and the tier is exposed on per-document API responses.

A single high-confidence document can constitute a signal. A cluster of low-reliability documents typically does not, regardless of volume. We do not equate loud with significant.

Inside the Trackers tab, SentiSense publishes two institution-ranking trackers built from SEC 13F filings: a top-10 concentration ranking and a largest-by-AUM ranking. This section explains what they measure and where the data is thin.

What gets ranked

The ranked universe is every institutional investor that files Form 13F with the SEC. Filing is mandatory for institutional investment managers with at least $100 million in US equity assets under management, due 45 days after each quarter end: Berkshire Hathaway, Vanguard, BlackRock, Citadel, university endowments, family offices, and roughly eight thousand others.

Each 13F discloses long positions in US-listed equities held at the quarter-end snapshot. That is the entire input. The 13F does not contain short positions, options, fixed income, commodities, FX, private holdings, leverage, cash balances, or fees. Both rankings are computed strictly on the long US-equity book a filer chose to disclose, so read them as a view of disclosed positioning, not of total fund size or performance.

What the two rankings measure

• Concentration ranks filers by the share of their disclosed book held in their top 10 positions: a proxy for conviction. A high share means a few names carry the portfolio; a low share means it is spread thin.
• AUM ranks filers by the total market value of their disclosed long US-equity book at the snapshot. This is disclosed-equity AUM, not a firm's total assets across all strategies and asset classes.

Coverage

Both rankings need a price for each holding at the quarter end. A filer whose book prices cleanly is a measurement; one with thin pricing coverage (non-listed securities, foreign listings outside our US-equity filter, or names delisted between filing and our run) should be read as a sample. We compute on the disclosed long US-equity book and nothing else.

Update cadence

New quarters are computed once the quarter is settled and the relevant 13Fs are filed (over the 45 days after quarter end); earlier filings refine the figure in place. Historical quarters are immutable unless we find a data-quality issue, in which case we re-run and note it in the change log.

The Hedge fund reported returns tracker records the net-of-fee returns large hedge funds state publicly, in the press and in their own investor communications. Rather than infer performance from disclosed 13F positions, every cell is a number a fund actually reported, with a citation you can open and read.

Where the numbers come from

These are not figures we compute. They are net-of-fee returns that funds, or reporters citing fund letters, have published: full-year annual returns and, where available, the latest year-to-date or interim figure. Each cell carries the primary source as a link and a short quote, so you can verify the number against its origin rather than trusting an aggregate. Coverage at launch spans roughly two dozen of the most-watched funds, with several years of annual history each.

How to read a cell

The headline column is each fund's latest reported annual return. Because funds report on their own schedule, the year that number refers to varies by fund, so the year is shown next to each value rather than in the column header: a fund that has reported 2025 shows 2025, one that has only reported through 2024 shows 2024. A blank cell means we have not found a citable public number for that fund and period, not that the return was zero. Some funds are deliberately opaque and may show only a single year.

The two alpha columns

Alongside the reported return, each fund carries two context columns. Alpha vs SPY is simply the fund's reported net return for a year minus the S&P 500 total return (dividends reinvested) for that same calendar year, so you can see the figure relative to just owning the index. 13F long-book alpha is a different, independent lens: the trailing one-year alpha of the firm's disclosed 13F long-equity book versus SPY, computed from its quarterly filings. It is deliberately one leg of the book, the long US-listed stocks only, not the fund's shorts, derivatives, fixed income, or private positions. For a market-neutral or macro fund the two columns can diverge sharply, and that gap is the point: it hints at how much of the story the long equity book alone tells. Read the 13F column as a cross-check, not as the fund's actual return.

What this tracker is, and is not

It is a faithful record of what funds have said publicly, net of fees. It is not an audited, standardized, or independently verified performance series: funds choose what to disclose and when, headline numbers can reflect a flagship share class rather than every vehicle, and selective reporting means strong years are more likely to be publicized than weak ones. Read it as sourced public claims, useful precisely because each one is citable, not as a like-for-like benchmark across managers.

Update cadence

New annual numbers are added as funds report them, which clusters in the first quarter of each year. We update existing cells in place if a fund restates or a better primary source appears, and we keep the original citation on every number so revisions are traceable.

This tracker reads public posts across a fixed allowlist of finance subreddits and builds a standing portfolio of the tradeable stocks the crowd has turned bullish on. A name enters when it is among the most-mentioned over a trailing window and sentiment leans bullish; its entry date and price are then frozen, and we score its return since entry against the S&P 500. The headline compares the equal-weighted portfolio to the index over the same windows, refreshed roughly monthly.

Read it as inferred positioning, not stated trades: mention volume and tone show what the crowd is talking about and how it leans, not what anyone actually bought. It is survivorship-free (entries stay, so weak picks are not dropped to flatter the average), skews toward large, liquid, heavily discussed names, and reflects the market regime. We exclude untradeable names and broad-market index funds, and omit any row we cannot price rather than show a fabricated return. Not advice.

This tracker ranks stocks by how the financial press is covering them right now: the ones the media is most positive on, and the ones it is most negative on. For each company we aggregate the sentiment of recent news coverage across our publisher set, weighting by source reliability the same way our per-stock sentiment does, and surface the extremes at both ends.

Read it as a measure of press tone, not of fundamentals or price: a "media darling" is a stock the coverage leans bullish on, which is not the same as a stock that will go up. It reflects the news flow of the moment and refreshes daily. Stocks need a minimum amount of recent coverage to appear, so thinly covered names are excluded rather than ranked on a handful of articles. Educational only.

The sentiment leaderboard ranks the stocks with the most bullish and most bearish aggregated SentiSense sentiment. Sentiment is our reliability-weighted mean of constituent document scores across news and social, expressed as a polarity in the range of minus one to plus one; the leaderboard simply sorts the universe by that value and shows both tails, with the driving story behind each name where we have one.

Only stocks clearing a minimum-mentions threshold are eligible, so a single article cannot push a name onto the board. It refreshes daily. As with all our sentiment surfaces, this describes the tone of available information, not a forecast of price. Educational only.

Where the leaderboard shows the current extremes, this tracker shows change: the stocks whose aggregated SentiSense sentiment moved the most over the past seven days, in both directions. The headline column is this week's sentiment minus last week's, so a name that swung from bearish to bullish ranks alongside one that was already positive and got more so.

Because it compares against a seven-day-old baseline, the board is empty until a full week of history has accumulated, then refreshes daily. The same minimum-mentions threshold applies at both ends of the window, so a stock has to be consistently discussed to register as a mover rather than appearing on a one-off spike. Educational only.

SentiSense publishes end-of-day options analytics for covered US stocks and for the ETFs we track, so the index and sector funds used as macro hedges (SPY, QQQ, GLD, TLT, the sector funds) carry the same readings as single names. Each night after the session closes, we process that session's full option chain for every covered name and reduce it to a compact set of daily readings. The data comes from licensed market data sources and is strictly end of day, not real time: every reading describes the most recent completed session.

Percentile-first

A raw implied-volatility level or put/call ratio means little on its own. A high-volatility tech name and a slow-moving utility live in different regimes, so comparing their chains cross-sectionally mostly measures which kind of stock each one is. Instead, every reading is ranked against that stock's own trailing history: put/call volume near the top of its own trailing range says the chain is unusual for that specific name, whatever the absolute number happens to be.

What we compute

• IV rank. Where at-the-money implied volatility sits within its own trailing range, on a 0 to 100 scale.
• Put/call percentiles. Today's put/call volume and open-interest ratios, each placed as a percentile of the stock's trailing window.
• 25-delta skew. The implied-volatility premium of 25-delta puts over 25-delta calls, expressed as a fraction on the same scale as IV itself (0.03 means 3 IV points), with its own percentile.
• Term structure. At-the-money IV at roughly 30, 60, and 90 days out.
• Open-interest walls and max pain. The heaviest strikes on the call and put side for a near expiry, and the strike at which the largest amount of option value would expire worthless.
• Unusually active contracts. Individual contracts whose session volume far exceeds their existing open interest, a description of fresh positioning in the chain.
• Options-implied sentiment. A composite in the range of minus one to plus one summarizing how the chain leans, derived from the chain alone, independent of our news and social sentiment.

Coverage and baselines

Coverage is a bounded universe of the most actively optioned US stocks, plus the ETFs we track, and it expands over time. The Radar board itself ranks stocks only; ETF readings live on each fund's own page. A covered name also needs history before percentiles mean anything: a name that has not yet accumulated enough sessions serves its raw readings without percentile context, and without the composite interest score, while its baseline builds. Treat a missing percentile as not enough history yet, not as a zero reading.

As with everything on this page, these readings are observational. They describe what the chain looks like today relative to its own past; they are not forecasts, price targets, or trade recommendations. A heavy open-interest wall is a description of where positioning sits, not a prediction of where price will go. Educational only.

The SentiSense Rating is a letter grade, A to F, that summarizes where a stock stands across seven independent signal families: crowd sentiment, smart money, options positioning, analysts, fundamentals, earnings and technicals. It is computed once per trading day for every stock SentiSense covers that has enough data.

Each rated stock gets a score from 0 to 100, carried to one decimal, and the letter is a band of that score. The score is built from a rank: how the stock's blended signals compare with every other stock rated that day, less a fixed adjustment for each risk condition the stock carries. Both constituents are published beside the score, so the grade never hides where it came from.

The Rating is a different object from the SentiSense Score. The Score is the continuous crowd-sentiment number for a stock and is one of the seven inputs here. The Rating is the composite grade. SentiSense publishes exactly these two headline measures.

How to read it

Score = rank percentile of the blended signals minus the sum of its risk adjustments, each worth up to 12 points, floor 10 when fewer than 6 of the seven dimensions are available. The score has exactly two constituents and both are published: the rank it started from, and every adjustment that took points off it, with what each one cost. The result is carried to one decimal and floored at 0.

The letter is a band of that score: A at 90 and above, B from 70 to 89.9, C from 30 to 69.9, D from 10 to 29.9, and F below 10. The letter and the number beside it are the same measurement, so they can never disagree.

The rank percentile is shown under the score, as the number the score started from. A stock reading 85.6 can be seen to have ranked in the 100th percentile before its risk adjustments, and the composition card spells the arithmetic out in one line: rank, adjustments, score, followed by each condition and what it cost.

Points come off the score rather than being applied to the letter, which is what an earlier draft of this method did. Limiting the letter piles every affected stock onto one band edge, where the grade stops separating them, while points keep them ordered. And the headline number stays the number the letter comes from, so a reader is never asked to reconcile a B with a 100th-percentile rank.

Most adjustments are graded rather than flat, because most of these conditions are matters of degree: a company just under the size threshold and one a hundredth its size are not carrying the same risk, and charging them the same would say they are. A flat charge also collapses the distribution, which is what the first production run showed. Four conditions stay flat because they are genuinely yes or no.

Because the letters are bands of the score rather than fixed slices of the market, their shares are not fixed. At most one stock in ten carries an A, since only the top tenth of ranks reaches 90 before any adjustment, and in practice far fewer do.

Under the letter, the composition card shows each of the seven dimensions on its own natural scale, its percentile among rated stocks, and whether it was available for this stock. Anomaly flags appear only when they are active.

A stock shows "Not rated" when too few dimensions have data, and a stock is never given a C by default.

The seven dimensions

Every dimension is turned into a percentile among the rated stocks that have it that day, so a reading from a bounded sentiment scale, a share count and a margin can be compared. The percentiles are then blended with fixed weights: crowd sentiment 16 percent, smart money 16 percent, options positioning 12 percent, analysts 12 percent, fundamentals 20 percent, earnings 8 percent and technicals 16 percent. The weights are engineering priors ordered by how persistent each signal is and by how much independent information each family carries: business quality from filed statements carries the most, the latest quarter's beat and its one-day reaction the least. They are not the output of a return backtest, and a later version may revise them.

Crowd sentiment enters through the SentiSense Score alone. Mention counts, social dominance and per-source slices are the same document stream and are not counted a second time.

Missing data and the coverage floor

A dimension with no data counts as the median, the 50th percentile, at its full weight. It is not dropped and the other weights are not rescaled. Removing the weight of a missing dimension would let a stock with two strong readings and five gaps outrank a stock that is strong across all seven, so thin coverage now pulls a grade toward the middle instead of lifting it to the top.

A stock is rated only when at least 4 of the seven dimensions are available and those dimensions carry at least 50 percent of the total weight. Otherwise it is Not rated for that day.

Coverage is also a risk condition, and a graded one, but it starts charging only once a stock drops below 6 of the seven dimensions: currently 6 points for each missing dimension past that point, so five dimensions cost 6 and four cost 12, while six or seven cost nothing. A stock missing one of the seven is well enough covered to grade on its rank alone. Below 6 the same condition also sets a floor, so a thinly covered stock never scores below 10 and is graded neither at the top of the market nor at the very bottom of it.

This is why smaller companies with little coverage, no options market and no analyst following are often Not rated rather than graded from one or two signals. A grade built from a single family would look like a grade and carry none of the information.

Anomaly flags

Three situations are flagged on the composition card when they are active. Two of them are also risk conditions and take points off the score while they are active: clustered insider selling, and institutional outflows against price. Unusual options flow is informational and costs nothing.

Risk conditions and what they cost

Eleven conditions can take points off a rank, each worth up to 12 points, and a stock carrying none of them scores exactly its rank. Seven are graded and scale with how far past the threshold the stock sits; four are flat because they are genuinely binary. They are listed in full because they are part of the definition of the score, not a separate test applied to it afterwards. The name in the first column is the value the API returns in the risk conditions array.

Condition thresholds are tuned as calibration data accumulates and may change without a version bump; the constituents, their points and the formula do not change without one.

Seasoning is counted in trading sessions rather than measured from a listing date, because there is no reliable listing date behind the grade. A stock that cannot show enough closes in the lookback window is either newly listed or too rarely traded to rank against the market.

Extended price is a condition rather than a ranked dimension because ranking every stock by its distance from its highs would reward the ones sitting at their lows, which is not what the measure is for. The technicals dimension ranks a stock's trend state against its own moving averages rather than against its highs, which is why the two coexist: one describes where the price sits in its own trend, the other prices how far it has run past that trend at the top of its range.

The conditions are why the top letter is scarcer than the top band. A recent listing with no filed statements and a spectacular rank gives back 24 points, and reads as a strong rank with real risk attached rather than as an A.

Freshness and cadence

The Rating is recomputed once per trading day, at 07:15 Eastern Time on the morning after each trading session, which puts the run on Tuesday through Saturday. It uses the prior session's crowd sentiment, the options reading computed that morning from the prior session's chain, the prior session's closing prices, and the latest available filings, consensus and earnings data. Institutional 13F positions are reported quarterly with a lag of up to 45 days after quarter end and are carried forward until the next report; congressional trades are disclosed with a lag of roughly 30 to 45 days.

Each daily Rating is stored, so the card can show how a stock's percentile has moved over the past 90 days. Because the grade is a rank, its history stays comparable even when the weights are revised.

What the Rating is not

The SentiSense Rating is not a recommendation to trade, a forecast of returns or a statement about whether a stock is suitable for anyone. SentiSense does not know your situation and does not give personalized financial advice. The Rating describes where a stock sits today across public signals, in a form that is transparent enough to check.

An A says the score is 90 or above: the blended signals rank near the top of the market and the stock carries few risk conditions. It does not say the stock will rise. An F says the score is below 10. It does not say the stock will fall. Signals can be stale, crowded or simply wrong, and every input has known blind spots listed below.

Known limitations

Fundamentals are ranked across all rated stocks, not within sectors, so structurally high-margin industries rank above structurally thin-margin ones regardless of quality within the industry. Institutional data lags by a quarter. Options positioning exists only for stocks with a liquid options market and enough history. Analyst coverage is thin for small companies. Crowd sentiment can be moved by a small number of loud accounts. The weights have not been validated against forward returns; a signal ledger that tracks each Rating against subsequent returns is planned, and its results will be published on this page.

Versions

Every stored Rating carries the methodology version it was computed under. The current version is 2026.09-v3. Changes to the formula, the weights, the floors, the letter bands, the list of risk conditions or the dimension definitions bump the version and are recorded here. Tuning an individual condition threshold does not.

2026.09-v3 (September 2026): seven dimensions, technicals included, blended at fundamentals 20 percent, crowd sentiment, smart money and technicals 16 percent each, options positioning and analysts 12 percent each, and earnings 8 percent, so business quality carries the most and the latest quarter the least. The EPS surprise leg is computed from the reported and estimated figures directly. A stock is rated at 4 of seven dimensions and thin coverage is charged below 6 of seven. Ratings stored under an earlier version keep it and are not recomputed.

In the API

The score, the letter, the rank percentile, the points taken off and the conditions behind them with what each one cost, the raw composite, the seven dimension percentiles and raw values, the flags and the methodology version are available through the SentiSense API for every rated stock, and the daily history is available as the sentisense_rating metric. The percentile, penaltyPoints and score fields are returned side by side so a caller can check the arithmetic, riskAdjustments pairs each condition with its points, and bucketLetter gives the band the rank alone would land in. See the API documentation.

Consumer Pulse tracks the products and services of the public companies we cover: how much each one is being talked about, how that volume changed against the week before, and how its parent stock moved over the same week. The table carries two groups, the products with the highest mention volume over the trailing 7 days and a second set ranked by week-over-week growth, with no product listed in both. Each row also carries the tone of the conversation around that product.

The question it asks comes from the social arbitrage approach Chris Camillo describes: watch what consumers are adopting and discussing, then ask whether the market has noticed yet.

What a mention is

A mention is one document in our corpus that resolves to a tracked product: a news article, a post or reply on X, a Reddit post or comment, a Substack piece, or a YouTube video. Each document counts once, on its publication day in New York time. Mentions are counted rather than weighted by audience, so a widely read article and a single post count the same.

App Store reviews are not mentions. A product with a mobile app carries two separate readings, a review count and an average rating, published beside the mention count and never summed into it. A wave of reviews and a wave of conversation are different events, and one number covering both would describe neither.

How a product qualifies

A product enters the table only when it has at least 30 mentions in the trailing 7 days. Below that a handful of documents can move a percentage by hundreds of points, which reads as a trend and is a sampling artifact.

Growth is the trailing 7 days of mentions minus the prior 7 days, divided by the prior 7 days, so a product that went from 100 mentions to 240 grew 140 percent. The prior week has to clear the same 30-mention floor, and the product has to carry at least 21 days of mention history, before any growth figure is published. Without both rules a product added last week compares its first full week against a near-empty one and posts a growth number in the thousands of percent, which measures when we started counting rather than anything happening in the world. A product that clears the volume floor but not these two rules is still listed on volume, with the growth cell left blank.

These figures are the current settings. They are tuned as we accumulate history and may change without a version bump; what a mention is, and the divergence rule below, do not change without one.

Unpriced attention

A row is marked "unpriced attention" when two things hold in the same week: mentions grew by at least 100 percent, so they at least doubled, while the parent stock's 7-day price return stayed within plus or minus 2 percent. In plain terms, the product got much louder and the stock did not move.

That is a divergence to research, and it is worth being exact about what it does not say. It is not a trade signal and does not predict the price. It does not say the attention is favorable, since a product can trend because customers are angry. It does not say the product is material to the company, and for a large company a single trending product is often a small share of revenue. The price return belongs to the parent company, not to the product. Attention and price also diverge when the market considered the news and concluded it did not matter, which is the ordinary case rather than the exception.

How products enter the tracker

Most tracked products are curated by hand: we record them in our knowledge base under the company that owns them, with the names people actually use. The rest arrive through discovery. When a phrase recurs across the documents we ingest and matches nothing we already track, it is put forward as a candidate. A grounding step then tries to attach it to a company we cover, using the sources it appeared in. A curator approves it, rejects it, or files it as a variant of a product we already track. Nothing is counted until it is approved, and the tracker publishes its most recent discoveries by name so the flow is visible.

Cadence and freshness

The search pass that actively looks for product conversation runs three times a week, on Monday, Wednesday and Friday mornings Eastern Time. The table itself is rebuilt more than once a day, on each run of the aggregate pipeline, from everything ingested up to that point, so a row can change during the day. Each snapshot carries the time of the run that produced it. The daily series behind each sparkline covers 30 days and is normalized against that product's own peak in the window, not against other products, so two sparklines are shapes to compare and not volumes.

Per-product history starts when the product enters the knowledge base, so a recently added product has a short series. That is a limit of our coverage, not a quiet week for the product.

What it is not

Consumer Pulse measures conversation, not sales, market share or revenue. Mention volume reflects what our sources publish and carries their bias: consumer software and large brands are discussed more than industrial products, and English-language sources dominate. A product whose owner is not a public company we cover ranks with the others, but it carries no stock or price column, so it can never carry the unpriced-attention flag. Nothing here is a recommendation to trade, a forecast of price, or personalized advice. Educational only.

Social Pulse compares an entity's average attention in a seven-day or 30-day window with the preceding equal-length window. Eligible measurement categories are normalized independently and then weighted equally, so a large count scale does not dominate a smaller one. Missing categories are omitted rather than treated as zero.

Coverage requires a current observation, at least 80 percent of dates in both comparison windows, and at least 14 observed days in the trailing 42 calendar days. Reviews and mentions also require at least five baseline events. Confidence reports history and category coverage, not the probability that a trend is true.

Anomalies use a robust z-score on log-transformed observations against the preceding 41 calendar days. A spike or fade requires a category z-score of at least 3.5 in absolute value, a same-direction category change of at least 20 percent, and a same-direction combined change of at least 20 percent. Attention measurements do not establish changes in sales, revenue, price, or investment returns.

• Small-cap and non-US coverage is thinner than large-cap US.
• Our models are trained predominantly on English-language content.
• Known model weaknesses include sarcasm, ironic framing, conservative forward guidance, and corporate euphemism.
• For time-sensitive signals, end-to-end latency from publication is constrained. Users building low-latency strategies should design for this bound.

Sentiment is not causation. A bearish shift describes the tone of available information; it does not predict price. We do not publish price targets or trade recommendations.

This methodology is reviewed semi-annually and on any material change to our data sources, models, or scoring. Changes are recorded in the change log. Prior versions remain accessible.

• No personalized advice. All published signals, reports, and scores are informational and educational. See the AI Disclaimer.
• No compensation from covered entities. We do not accept payment in exchange for coverage, rating, or placement.
• Team members may trade. Members of our team actively participate in the markets and use SentiSense as a research tool. Published research is generated by AI systems built for balanced analysis and does not reflect any team member's personal positions.
• Data licensing. Commercial licensing relationships do not influence the content or scoring of our published research.