Methodology

Who we follow

We follow a short, hand-picked list of 3 shows and people who talk regularly and seriously about markets, AI and crypto. They're chosen for track record, independence and range: the list deliberately includes people who disagree with each other. A voice is added or removed by a person, never by an algorithm, and every voice page links to the original channel or feed.

From episode to summary

1. Discovery. Each day we check every voice's YouTube channel or podcast feed for new episodes. Shorts, clips and trailers are skipped.
2. Transcript. We use the video's captions, the publisher's own transcript where the podcast feed provides one, or a machine transcription of the audio. Transcripts are used for analysis only and are never published.
3. Extraction. A language model reads the transcript and records, for each of markets, AI and crypto, the episode's stance on a five-point scale (strongly bearish -2 to strongly bullish +2), conviction and time horizon, plus the specific claims and predictions made, with the time each one was said.
4. Checks in code. Before anything is saved, every quote is compared with the transcript and dropped unless it is word-for-word and 25 words or fewer; timestamps outside the episode are dropped; stances are clamped to the scale.

The consensus index

First, each voice gets a stance per topic: the average of the stances in all its episodes from the last 45 days, weighted by conviction (low 0.5, medium 1, high 1.5), by how much of the episode was about markets, AI or crypto, and by recency, halving every 14 days. One off-topic guest can't flip a show's view, but last week counts for more than last month. Then the voices are averaged, each counting once and weighted by how recent its latest episode is. The result runs from -2 to +2 and is shown as an index from -100 (risk-off) to +100 (risk-on). Labels: risk-on from +50, leaning risk-on from +18, mixed between -18 and +18, leaning risk-off down to -50, risk-off below that. Agreement is the weighted spread of stances. All of this is arithmetic in code; the model never produces a number we publish.

Each voice counts once however often it publishes, so a daily show doesn't outvote a weekly one.

The written summaries

The short prose summaries are written by a language model from the extracted views and claims only. Every sentence must cite the episodes it draws on; code deletes any sentence that cites nothing, cites an episode that wasn't in its input, or states a number that doesn't appear in the episodes it cites. Citations are the numbered links after each sentence.

Weekly editions

Each Monday the previous Monday to Sunday is frozen into a weekly edition. Editions are never edited afterwards, so they are a fair record of what was said at the time.

What gets a page

We only ask search engines to index pages with enough substance: episodes with at least 1500 words of transcript and several distinct claims, weekly editions covering at least 3 voices, and topic pages for subjects at least 3 voices discussed in the last 30 days.

Known limitations

• Models misread sarcasm, hypotheticals and devil's-advocate arguments. Follow the timestamp links before relying on any claim.
• Machine transcripts mishear names and numbers. Quotes are checked against the transcript, not the audio.
• Shows with several hosts are scored as one voice, so a disagreement between co-hosts can average out to "mixed".
• The consensus reflects only the voices we follow, not the market or investors in general.

Spotted a mistake? Email hello@readonrisk.com with the page link.