Technical Methodology

Every 6 hours we poll the YouTube Data API for new uploads across our curated channel list. Each channel has a stable channel_id and an admin-tuned reliability weight (1–100) that reflects the creator's track record.

New videos land in the videos table with metadata only — transcript and mention extraction run as separate stages so a failure in one doesn't block the next channel.

Two-pass strategy, every 5 minutes, batched per video:

1. Pass 1 — native captions. YouTube auto-captions (or creator-provided) via youtube-transcript. Free, fast, ~80% coverage.
2. Pass 2 — AI speech-to-text. For the ~20% with no captions, Google Gemini transcribes the audio directly. Slower and metered, but no video is left behind.

Both paths produce a single transcript column with the source tagged (youtube / ai / none) so we can audit quality later.

Each transcript is passed to Gemini with a structured-output schema. The model returns a JSON array of mentions, each carrying:

{
  "ticker": "NVDA",
  "stance": "bullish" | "neutral" | "bearish",
  "confidence": 0.0 - 1.0,
  "timestamp_seconds": 412,
  "excerpt": "...verbatim quote..."
}

Stance reflects what the creator said about the ticker — not whether we agree. Confidence is the model's certainty that the ticker was actually being discussed (vs. a passing reference or a misheard word).

Models hallucinate. Before any mention can move forward, the ticker symbol is checked against Yahoo Finance — if the symbol doesn't resolve to a real listed instrument it's quarantined and removed by a weekly cleanup-invalid-tickers cron job. Validated symbols are cached in the tickers table with exchange, name, and logo.

This is the real math, run on every recompute (auto-chained after every new transcript batch). For each ticker T across a rolling 14-day window:

sign = +1   if stance == "bullish"
       -1   if stance == "bearish"
       +0.2 if stance == "neutral"

mention_value = sign × confidence

for each creator C who mentioned T:
  creator_avg[C] = mean(mention_value over C's mentions of T)
  creator_avg[C] = clamp(creator_avg[C], -1, +1)

raw = Σ ( weight[C] × creator_avg[C] )   for all creators C of T

theoretical_max = max_weight × number_of_creators
consensus_bonus = 1.15 if creator_count >= 3 else 1.00

deviation = (raw / theoretical_max) × 50 × consensus_bonus
score     = round( clamp( 50 + deviation, 0, 100 ) )

A score alone doesn't get published. The pick must clear:

• ≥ 2 distinct creators, OR
• ≥ 3 total mentions from the same creator

This kills single-mention noise. A ticker namedropped once by one creator never becomes a published recommendation, regardless of how confidently the model extracted it.

The previously-published score remains live until the next recompute changes it by ≥ 5 points or the contributing mentions change — at which point the AI thesis is invalidated and regenerated.

Every published pick carries a written thesis — the "why" you see on the stock page. It's generated by a separate Gemini pass that ingests:

• The ticker's verbatim mention excerpts (with timestamps)
• Each contributing creator's name and weight
• The final score and stance distribution

When the underlying mentions or score shift materially, the thesis is marked stale and re-queued. A worker drains the queue every 5 minutes, so the public site never shows a thesis that contradicts the current score.

We measure ourselves. Every published pick with score ≥ 70 is auto-enrolled into a hypothetical $100 position, captured at the next trading day's open after publication.

entry_date  = first trading day after published_at
entry_price = open[entry_date]
spy_entry   = open[entry_date] for SPY

window prices captured at +1M, +3M, +6M, +1Y:
  pick_return  = (price[window] / entry_price) - 1
  spy_return   = (spy[window]   / spy_entry)   - 1
  alpha        = pick_return - spy_return

Snapshots fill in automatically as windows mature (daily cron). We aggregate per creator and site-wide to a leaderboard. The dashboard is currently admin-only until the sample size is large enough to publish honestly — back-filling old picks with synthetic entry prices would distort the record, so we're letting it accumulate forward-only.

Assumes no fees, slippage, taxes, or dividends. α is purely illustrative — past performance doesn't guarantee future results.

What we explicitly do not do:

• No price targets. We don't predict where a stock will go — we surface what high-conviction creators are saying, weighted by their reliability.
• No recency decay yet. Within the 14-day window all mentions count equally. A future revision may weight more recent mentions higher.
• No sector or macro weighting. A 100 score on a microcap means the same thing as a 100 score on a megacap. Position-sizing is on you.
• No short signals emphasis. Bearish stances are tracked symmetrically but the audience reality is that most viewers act on long ideas.
• Not investment advice. This is a signal aggregator. Read the disclaimer.