Methodology
How InfluAlpha classifies calls, prices them, calculates returns and ranks tracked channels, including data limitations.
Understand these figures or report an error: Methodology Corrections policy
What counts as a recommendation
New extractions require the speaker’s own explicit current buy or sell recommendation. Neutral mentions, optimism, holdings or past trades alone, hypothetical or conditional future scenarios, and quoted opinions without explicit current endorsement are excluded. Quotes and surrounding sentences must be verbatim and source-matched before storage. Catalog matching and contradiction filtering still apply. Historical classifications have not been reprocessed. AI classification and mechanical source matching are not human verification. Call links cite existing anchors; timestamp links require a unique exact quote match and a retained source segment offset. Otherwise the video link is untimed. Creator responses and disputes do not establish a completed correction; see the corrections policy.
Entry prices and latest prices
A video call requests a price for the video’s publication date, falling back to the ingestion date when publication is missing. It does not use the spoken timestamp, intraday execution or the next tradable price after publication. For equities, the lookup walks back over weekends, recognized holidays and missing data, trying up to five candidate trading dates. FMP supplies US prices; brapi supplies B3 prices; CoinGecko supplies crypto prices in USD. Unknown exchanges fall back to FMP. The application keeps the latest price’s returned date, but does not retain the actual lookup date for a recommendation or benchmark entry. A displayed publication date therefore does not establish the exact entry-price session.
Price fallback limitations
FMP uses the requested date’s bar if found, otherwise the first returned bar. B3 can fall back to a current quote for requests within seven days, or within 30 days when the historical array is empty. Crypto requests for today or yesterday use the current-price endpoint; older dates use historical snapshots. Crypto has no exchange closing bell. These fallbacks can make an entry differ from the price on the requested day. Each asset and benchmark has its own latest snapshot; their latest dates are not forced to match, and a site-wide freshness date does not mean every asset is equally fresh.
Returns and holding periods
Let E be the entry price and L the latest stored price. Buy return (%) = 100 × (L − E) / E. Sell return (%) = 100 × (E − L) / E. These measure directional price changes from entry to the latest snapshot, with no fixed exit date, annualization or closure when another recommendation appears. A “12-month” channel figure selects videos published in the last 12 months; it does not give every call a 12-month holding period. Fees, spreads, taxes, borrowing costs, leverage, position sizes and currency conversion are not included. A sell score is a hypothetical directional result, not a verified short trade.
Benchmarks and excess return
NASDAQ and NYSE calls use SPY (an S&P 500 ETF); B3 uses BOVA11 (an Ibovespa fund); crypto uses BTC. Other exchanges have no defined benchmark. Benchmark return (%) = 100 × (latest benchmark price − benchmark entry price) / benchmark entry price, always long, including for sell calls. Excess return, labeled alpha, is the call return minus the benchmark return, in percentage points. This is simple excess price return, not risk-adjusted alpha. Benchmark entries request the video’s entry date, subject to the same lookup limitations. Missing benchmark prices leave alpha unavailable. Calls on their own benchmark, such as BTC against BTC, have no alpha.
Averages and weighting
For each video, available returns are deduplicated by asset and action, then equally averaged. Alpha is averaged separately over calls with an available alpha; each stored video figure is rounded to two decimals. Channel return is the equal average of stored video returns for visible, analyzed videos published in the last 12 months. Channel alpha averages available video alphas within that set. Channel figures are rounded to two decimals. Thus a video with one call has the same weight as a video with ten; this is not a capital-weighted portfolio. Profile asset tables instead average individual scored call rows across published videos, including older videos. The asset 12-month metric averages priced buy-call rows only; it does not deduplicate or apply the call table’s age filter.
Ranking eligibility and win rate
The top-influencer board requires at least five visible, analyzed videos with a stored return in the last 12 months, at least ten priced call rows in the stored channel count, and a non-missing channel return. It shows up to ten channels, ordered by alpha, then raw return, then channel name; missing alpha sorts last. The channel directory is a sortable roster and includes ineligible channels. Channel win rate = positive-return call rows / all call rows with positive entry and latest prices in those scored videos × 100. Flat calls are in the denominator but are not wins. This count does not deduplicate repeated calls or apply the call table’s age filter. Board call totals count all recommendation rows in its scored videos, so they can differ from the win-rate denominator. Winning means positive raw return, not beating the benchmark.
Missing, early and stale scores
Missing prices or a zero entry make the return unavailable. Missing alpha is omitted from alpha averages, not counted as zero. Call tables suppress video calls when the asset’s latest price date is at most one calendar day after publication; this is a calendar-day test, not a full exchange-session count, and is skipped if either date is missing. Stored video scores do not apply that age test, and their arithmetic does not require positive prices as the channel win-rate query does. These differences can produce different counts and averages across pages. If no video returns can be calculated, the scoring job leaves an existing stored score untouched. Missing or stale data is not evidence of zero performance.
Market holidays
Equity lookups move backward rather than forward. The US calendar includes fixed holidays and Martin Luther King Jr. Day, Presidents’ Day, Memorial Day, Labor Day and Thanksgiving, but does not fully cover observed holiday shifts, Good Friday or exceptional closures. B3 includes fixed holidays but omits movable holidays such as Carnival, Good Friday and Corpus Christi. Provider walk-back can compensate for missing bars, but this is not a complete historical market calendar. Latest equity targets use a settlement buffer: 17:00 New York time and 19:00 São Paulo time; crypto is requested daily.
Splits and dividends
The code reads provider close fields and does not separately add dividends or reinvest distributions. These are not verified total returns; the application does not establish a uniform provider adjustment basis. A weekly split check probes up to three of an asset’s oldest priced video calls within the last 12 months. If stored price / freshly fetched price falls outside 0.75–1.25, it refetches that asset’s entries in the window, rewrites those outside the band and queues video recalculation. This is a drift heuristic, not a corporate-action ledger. It can miss smaller changes, fail when history is unavailable, or react to vendor revisions. It does not repair older calls, user-call entries or benchmark-entry fields.
Refresh behavior
Production configuration schedules new channel analysis daily at 06:00 UTC and transcript batch submission every two minutes. Recent asset prices are scheduled at 07:00 Tuesday–Saturday, with crypto-only runs on Sunday and Monday and explicit benchmark refreshes. Video scores run daily at 09:00, asset metrics at 10:00 and channel metrics at 11:00; portfolio rebuilding and index points follow at 12:00 and 13:00. Split checks are scheduled Saturday at 05:00. Video refresh covers visible analyzed videos between one day and 12 months old, using ingestion date if publication is missing. Missing video-call entries retry for seven days from recommendation creation; an administrator can request a backfill later. Existing entries are ordinarily retained. Queue delays, failures and provider availability can leave stages out of sync; these schedules are not delivery or freshness guarantees. The scheduled price universe is assets with video or user calls in the last 12 months plus benchmarks. Opening an asset page can also queue a stale-price refresh; completed extraction schedules video scoring after five minutes. These are asynchronous requests, not immediate updates.
Derived portfolios and trader profiles
The derived channel portfolio groups priced calls from published videos in the last 12 months by asset and action and averages their entry prices, rounded to two decimals. Mention counts are not share allocations. Portfolio summary returns equally weight available position returns; its “in profit” share includes flat positions. The index is 100 × (1 + equal average position return as a fraction), recomputed against the current book, not a compounded investable track record. Separately, user calls are manually logged, dated on submission and priced asynchronously from the asset snapshot after a stale-price refresh attempt. Their entry date is retained, but delayed pricing can use a later snapshot. They become scored only after a newer asset price date. Public traders rank after five scored calls, by alpha, return and display name, across their recorded calls without a 12-month cutoff. Their averages and win rate weight scored call rows equally. Users can delete their calls; the record is not immutable.
Worked example — hypothetical, not a real creator
Assume two distinct calls in one video, all prices available and old enough to score. A buy enters at 100 and is now 120: 100 × (120 − 100) / 100 = +20%. A sell enters at 100 and is now 90: 100 × (100 − 90) / 100 = +10%. Suppose each benchmark moves from 200 to 210: +5%. The call alphas are +15 and +5 percentage points. Video return = (20 + 10) / 2 = 15%; video alpha = (15 + 5) / 2 = 10 percentage points. Both calls have positive raw returns, so win rate = 2 / 2 = 100%. If a second eligible video has one call with −5% return and −8 percentage points alpha, the two-video channel averages are 5% return and 1 percentage point alpha; call win rate is 2 / 3 ≈ 66.67%. Three calls in two videos do not meet board eligibility.
Coverage and interpretation
The record contains selected tracked channels, available transcripts, supported catalog assets and successful analyses, not every creator, video, statement or tradable asset. Failed, hidden and no-recommendation analyses are excluded from public scoring. Missing histories, delistings, ticker ambiguity, transcription errors and extraction choices can bias the sample. Channel attribution is not verified identity or endorsement; a channel claim verifies control through the application’s claim process. No complete coverage, human review of every call, execution at the stated price or future result is guaranteed. Use the source video and the correction options when something looks wrong.