By: Max
How to Tell Which KOLs Are Providing Alpha and Who’s Selling a Story
I’ve been in crypto long enough to lose count of the cycles and slogans. The one that stuck, “this is an attention market”, is both true and useless. We all say it, but we rarely measure it. I’ve been obsessed with changing that: not by arguing about who’s influential, but by quantifying it. Clip by clip, post by post, trade by trade.
Treat creators like time-series data. If persuasion lives on camera, then we should be able to score it. Segment the exact moment a call is made, extract the trade, measure the conviction, and settle it against real benchmarks. Multimodal AI becomes a truth filter. Attention in, receipts out.
The VideoConviction Proof
A new benchmark quietly proves the point. Researchers took long-form YouTube “finfluencer” videos and did what quants do to prices: segment them, annotate them, test them. Instead of treating a 10-minute vlog as one blob of content, they isolate the exact window where a trade is made and score the delivery: ticker, action (buy/sell/hold), and a 1–3 conviction rating based on tone, facial cues, and title-content alignment.
Across 288 videos, that produced 687 recommendation segments and 6,000+ expert labels. Enough to backtest outcomes.
Two findings matter. First, video as video matters: letting models “see” frames improves fact extraction (tickers appear on slides and charts), whilst transcripts alone miss visual tells. Second, the hard part isn’t reading a ticker, it’s deciding whether the creator actually issued a trade and how strongly. Even with video, models often confuse commentary with recommendation. The unit of analysis must be the segment where the call happens, not the whole episode.
Then comes the punchline: tested against markets. When researchers ran simple strategies since 2018 (six-month holds, no look-ahead) and compared them to QQQ and SPY, the index funds won on risk-adjusted terms. A contrarian “Inverse YouTuber” basket topped the S&P on annual return, but with a worse Sharpe. High-conviction picks beat low-conviction ones and still lagged QQQ.
In plain English: the median recommendation underperforms. Confidence sells better than it compounds.
What You Actually Build
Start where the medium lives: video first. Ingest the clip and its metadata, find the trade moment, align transcript to frames so voice, text overlays, and facial affect stay in sync. Extract the instrument and action with rules that punish vague phrasing (“I’m watching SOL” ≠ “I’m buying SOL”). Score conviction from delivery, not just adjectives.
Then price the claim. Normalise to point-in-time trades, hold for a defined window, and report P&L and Sharpe versus the appropriate benchmark: S&P 500 for equities, BTC/ETH beta baskets or sector indices for crypto. The same logic that sorted stock YouTube ports neatly to crypto YouTube, TikTok, and X Spaces, where hype cycles are shorter and thumbnails louder.
But don’t stop at video. Scrape their X accounts too. Use vision AI to read the screenshots they post (broker fills, charts with tickers tucked in the corner, “I’m buying” slides) and align those images to the exact claim in time. Pair that with transcripts and on-screen text so we’re not guessing what was said or when.
This is the editorial test markets deserve:
- See the moment the trade is made (segments beat full videos)
- Hear the certainty (conviction, scored multimodally)
- Extract what’s tradable (ticker + action, not vibes)
- Backtest against a benchmark with real risk maths
Do that, and “who to follow” stops being a popularity contest. You get Noise-Adjusted Influence: a creator’s risk-adjusted track record, calibrated by how well their expressed conviction maps to realised outcomes. Your first filter is the video, but your final arbiter is the portfolio.
The Scoreboard
One profile per influencer. One methodology. Every creator’s calls timestamped from video and posts, tracked against real benchmarks. Same horizons, same rules, no wiggle room. Show P&L, Sharpe versus benchmark, hit rate by action, and a small “follow” or “fade” toggle that spins a live basket with liquidity filters.
Segment-level receipts you can scrub to. A conviction calibration curve that shows whether their “strong buy” historically beats their own “maybe”. For crypto, layer beta-controls and sector views (L1s, DeFi, AI, memes) so you can see who rides beta and who actually generates alpha.
It’s not about dunking on hype. It’s about standardising truth. Put every claim on the same field, settle it in P&L, and let the benchmarks do the talking. The real ones surface. The noise fades.
Be the first to comment