Experiment: AI channel with trading signals from news

Table of contents

• Goal
• Experiment setup
• Example signals
• Resulting channel
• How to build a similar channel
• Ideas and improvements

Goal

The goal is to demonstrate the service capabilities and show how to reproduce the experiment — not to assess LLM prediction accuracy. Still, I’ll try to review a few examples from the channel.

I also hope some readers will find clever ways to apply AI.

Experiment setup

• Task: Connect a set of sources (Telegram and X/Twitter channels about finance and crypto), monitor news, and produce structured trading signals — with explanation and analysis.
• Sources: I selected public Telegram channels that regularly publish news and analytics on crypto and stocks. (The full list is not provided here — this is an internal selection for the experiment.)

Mechanics:

1. The bot reads posts/news from connected channels.
2. Using a prompt, an LLM (gemini-2.5-flash) analyzes the content (event, sentiment, mentioned assets).
3. The bot generates a structured signal and publishes it to the target channel.
4. There’s a basic dedupe filter and autoposting interval enabled.

What I wanted to check:

• Do meaningful signals emerge from the news stream?
• How well and consistently does the model analyze news?

Prompt used:

Your role is a professional manager of a Telegram channel with trading signals. You receive news posts from other sources.
Your task is to write a new post — a trading signal. The post must include what to trade, a brief news summary, how the news affects the market, analysis, a clear recommendation, and a disclaimer.

Post format:
1. <b>[Token, stock, company name]</b>
2. News: <blockquote>[short news summary]</blockquote>
3. Impact: <b>[low, medium, high]</b>
4. AI analysis: <blockquote>[brief analysis and reasoning]</blockquote>
5. AI recommendation: <b>[buy, sell]</b>
6. <i>This AI signal is not financial advice. It’s generated automatically and does not guarantee profit.</i>

Allowed HTML tags: <b>, <strong>, <i>, <em>, <code>, <s>, <strike>, <del>, <u>, <pre>, <blockquote>, <tg-spoiler>. Do not use other tags or markdown.
 
Filtering. If the incoming news contains any of the following, respond with "no signal" only:
- Advertising (promotions, third-party services). Any invitations to external resources.
- Memes, jokes
- Casual discussions, chatter
- Not related to the market or a specific token/company
- News with impact below medium

Inside @signaller_ai_bot, the user prompt is concatenated with the raw text and image of the original post.

Example signals

Here are a few signals we received.

I’m not a professional analyst, but I’ll try to compare the price movement on an hourly chart after the signal.

1. Signal to buy BTC:
Buy BTC 23.08.25 23:34 +3GMT BTC hourly chart on Coinmarketcap

The bot rated the impact as high, but the chart doesn’t show a strong reaction.

2. Signal to buy altcoins:
Buy altcoins 24.08.25 02:42 +3GMT CMC Altcoin Season Index (Coinmarketcap)

It looks like an altcoin season forming. Hard to say how it will play out overall.

3. Signal to sell PIPPÍN:
Sell PIPPÍN 24.08.25 06:03 +3GMT PINPIN hourly chart (Coinmarketcap)

The signal arrived during a large red candle. The recommendation looks reasonable here.

4. Signal to buy AAPL:
Buy AAPL 24.08.25 17:17 +3GMT AAPL hourly (TradingView)

The signal came on Sunday evening when the market was closed. Still, the first candle after open was quite strong.

5. Signal to buy ETH:
Buy ETH 24.08.25 21:34 +3GMT ETH hourly (Coinmarketcap)

This buy signal arrived at the peak. Although the model rated the impact as high, buying there would likely lead to a loss.

6. Signal to sell XTERUSDT:
Sell XTERUSDT 25.08.25 00:26 +3GMT XTER/USD hourly (Coinmarketcap) The sell signal came at the bottom. The overall trend looked bearish, but the news didn’t seem to matter at that exact moment.

7. Signal to buy LINK:
Buy LINK 26.08.25 13:45 +3GMT LINK hourly (Coinmarketcap) The impact was rated high and the chart looked like a reversal forming. Potentially a good one.

Resulting signals channel

You can explore the channel here:

How to build a similar channel

Preparation

1. Find the bot: @signaller_ai_bot. On first start, it will suggest joining a channel to get bonus tokens — follow the instructions.
2. Gather sources to monitor. Pick 5–20 relevant channels posting news, analytics, or press releases. Prefer verifiable sources (official announcements, reputable media). Use /add to add them.
3. Set /prompt with the format and filtering rules. Use the special phrase “no signal” — if the LLM outputs it, the post won’t be published.
4. Telegram posts are ingested instantly and usually processed within 5 seconds. X (Twitter) posts are checked every 15 minutes.

Creating and connecting the channel

1. You can receive signals in your chat with the bot without auto-posting — connecting a channel is optional.
2. Create a Telegram channel and copy its username (e.g. @my_signals_channel).
3. Add @signaller_ai_bot as an admin. Grant posting permissions.
4. Link the channel to the bot. Send /channel to the bot with your channel link or username.
5. Configure /settings: set post interval (e.g. max 1 post per 30 minutes), enable dedupe, and switch on auto-posting.

Moderation and responsibility

Ideas and improvements

• You can use other available models. Get the list via /models.
• For teamwork, add the bot to a group chat. Billing will be tied to the user who added the bot first.
• Use high-quality sources; many original newsbreakers are on X (Twitter).
• Iterate your prompt. Ask AI explicitly for what you need.