Editorial policy

How a Recommendation Gets From YouTube to Yatırım Radar

Yatırım Radar does not write commentary — it structures the commentary that Turkish YouTube finance speakers are already making in public. This page explains the full technical and editorial pipeline: where we get our data, how AI assists, what the editors check, and what we will not publish.

Yatırım Radar Editorial TeamFebruary 1, 2026April 21, 2026AI-drafted summary reviewed and verified by the editorial team.

Why publish a methodology?

Turkey's retail investor base is growing faster than its financial-literacy infrastructure. Thousands of hours of expert commentary are posted to YouTube every week, most of it never indexed in a searchable way, and much of it flagged months later when a call turned out wrong. We built this platform to solve one specific part of that problem: give individual investors a way to see what each expert actually said, about which asset, on which date, with a link to the original video — fast.

But "structured expert commentary" only has value if readers understand exactly how it was structured. AI systems can hallucinate. Transcripts can be misheard. Speakers can be quoted out of context. So we publish, in detail, what we do to minimize each of those risks — and what we still cannot guarantee.

The end-to-end pipeline

1. Source selection

We maintain a curated list of Turkish YouTube finance channels. Each channel is scored by editorial staff on criteria like consistency, disclosure practices, track record transparency, and adherence to Capital Markets Board (SPK) disclosure rules. Channels can be added, paused, or removed — and every change is logged.

2. Automated ingestion

A Python service (yt-dlp + youtube-transcript-api + MLX Whisper) watches the channel list every 20 minutes. When a new video is detected, we fetch the audio, attempt to read the uploader's native captions first, then fall back to MLX Whisper speech-to-text when captions are absent or low quality.

3. LLM analysis with grounded prompts

The transcript is sent to Google Gemini with a domain-specific system prompt. The prompt forces structured JSON output: asset symbol, direction (5-level), reasoning, SPK-safe summary, detailed analysis, and a confidence score. The model is instructed to refuse fabrication — if the speaker did not make an actionable call, the video is classified as non-recommendation.

4. Validation + normalization

Before anything is written to the database, we normalize asset symbols against reference tables (varliklar, konusmacilar), verify the direction value falls within the 5-level enum, and run a confidence gate. Low-confidence outputs are flagged for editor review instead of being auto-published.

5. Editorial review

Each new recommendation appears in an internal admin queue. Editors verify the asset symbol is correct, the direction is consistent with the speaker's actual claim, and the SPK-safe summary does not mislead readers. Rejected items are sent back with a reason so the prompt can be tuned; accepted items move to publication.

6. Delay policy + publication

Some channels ask us to honour a publication delay (stored per-channel in analysis_delay_hours). Free-tier users see the recommendation after that delay has elapsed. Premium subscribers, when subscribing through the mobile app, receive content as soon as it clears review. Every published card links back to the exact YouTube video and timestamp.

Why we store two versions of every summary

The Capital Markets Board of Turkey (SPK) regulates who may give investment advice. Yatırım Radar is not licensed to give advice, and much of our data comes from speakers who are also not licensed. To stay on the right side of that line, every analysis we publish is stored in two forms:

Detaylar_SPK_Onayli_* (TR/EN)

A neutral, third-person paraphrase of what the speaker claimed. No imperatives ("buy", "sell"), no prediction language framed as advice. This field is visible to free-tier users and constitutes our default rendering.

Detaylar_Net_Tavsiye_* (TR/EN)

The speaker's actual framing, including imperative language where they used it. This expanded summary is gated behind a Premium subscription and behind an explicit disclosure: "this is the speaker's language, not our recommendation".

The 5-level direction system

Investor commentary is not binary. A speaker who says "wait and see" is not the same as one who says "full exit". To preserve that distinction, we map every call into five buckets. Free-tier users see a collapsed 3-level view (Buy / Hold / Sell); Premium subscribers see all five levels.

Level	Label	Meaning
4	Strong Buy	Speaker expresses conviction — typically a "top pick" framing with sized allocation guidance.
2	Buy	Speaker recommends accumulating, though without conviction-level framing.
1	Hold	Speaker is neutral: "wait and see", "already positioned — not adding".
0	Sell	Speaker recommends reducing or exiting a position.
3	Strong Sell	Speaker warns readers off the asset entirely — "avoid", "exit fully".

Delay policy and non-recommendation filter

Not every YouTube finance video produces a recommendation. Many are macro commentary, news recaps, or personal-finance explainers — valuable content, but not a "call" on any specific asset. Our system explicitly classifies these as non-recommendation videos and stores their identifiers in the processed_non_recommendation_videos table. We do this for two reasons: we never want to fabricate a call, and we do not want to re-process the same non-recommendation video on the next polling cycle.

Some channel owners explicitly ask us to delay publication so their paying audience gets the signal first. We honour this with a per-channel field called analysis_delay_hours. During that delay, free-tier users see a clearly labelled "released in X hours" card instead of the substantive content, and no advertising is served next to the gated card. After the delay expires, the content becomes visible to everyone.

What we will not do

We will not fabricate calls. If a speaker did not make an actionable recommendation, the video is classified as non-recommendation and no row is written to yatirim_onerileri.
We will not publish paid placements as if they were independent expert opinion. If a speaker is running a sponsored segment, either the editorial reviewer rejects the item, or the card explicitly labels it as sponsored content.
We will not target or coordinate promotion of specific assets. Our editorial policy prohibits editors from surfacing a recommendation for personal gain, and our infrastructure does not support ordering cards by anything other than recency, category, and speaker filters that the reader chose.
We will not remove unfavourable historical recommendations. A speaker's past track record — including wrong calls — is part of what makes this platform useful. We correct factual errors (wrong symbol, wrong direction) when reported, but we do not delete items simply because the market moved against them.

Correction and feedback policy

If you spot a misattributed quote, a wrong ticker, or a call framed more aggressively than the speaker actually made it, please tell us. The fastest route is the contact form, but a direct email to [email protected] also works and gets human-reviewed within one business day.

When we correct an item, we update the dateModified field on the page and leave a visible "Son güncelleme" / "Last updated" byline. We do not silently rewrite history.

Report a correction Masthead

This methodology is a living document. Last substantive update: 2026-04-21. When we change how recommendations are produced, we update this page and note the change on the relevant blog post in Yatırım Radar's editorial feed.