Social media is full of noise. For every actionable trading signal, there are hundreds of irrelevant posts, jokes, and false positives. TradeFollow's AI is designed to separate the signal from the noise—understanding context, intent, and relevance to surface only what matters for trading.
This article explains how our signal filtering works.
The Challenge: Signal vs. Noise
The Volume Problem
A single active crypto account might post: - 20+ tweets per day - Mix of personal, promotional, and trading content - Retweets, replies, and original content - Varying levels of seriousness
Without filtering: You'd receive dozens of irrelevant alerts daily per account.
With intelligent filtering: You receive only the actionable signals.
Examples of Noise vs. Signal
| Post Content | Classification | Why |
|---|---|---|
| "Just bought some BTC, feeling bullish" | Signal | Clear action (bought) + asset (BTC) |
| "Bitcoin is the future of money" | Noise | Opinion, no actionable info |
| "Binance listing $XYZ tomorrow" | Signal | Specific event, verifiable, tradeable |
| "Can't believe how dumb crypto bros are" | Noise | Commentary, no trading relevance |
| "Sold my entire ETH position at $4000" | Signal | Clear action + price point |
| "GM crypto fam! Let's have a great day" | Noise | Social, no trading info |
How Our AI Filtering Works
Layer 1: Keyword Detection
The first filter identifies posts that might be trading-related:
Positive Keywords: - Action words: buy, sell, long, short, accumulating - Asset mentions: token symbols, project names - Event words: listing, announcement, launch, airdrop - Sentiment indicators: bullish, bearish, moon, dump
Negative Keywords (reduce priority): - Humor indicators: lol, lmao, joke - Question formats: "should I buy?" - Hypotheticals: "if I had," "would be nice"
Layer 2: Natural Language Understanding
Keywords alone create too many false positives. Our NLP models understand:
Intent Recognition: - Is this a statement of action taken? - Is this a prediction or recommendation? - Is this a question or hypothetical? - Is this sarcasm or serious?
Example:
"I would never buy BTC at these prices"
Keywords: buy, BTC ✓
Intent: Negative/rejection ✗
Classification: NOT a buy signal
Layer 3: Context Analysis
Posts don't exist in isolation. We analyze:
Thread Context: - Is this a reply to something? - What's the conversation about? - Does context change the meaning?
Account Context: - Does this account usually share trading info? - Is this consistent with their normal behavior? - What's their track record?
Market Context: - Is the mentioned asset actually tradeable? - Is the market open/liquid? - Does this align with market conditions?
"To the moon!" means something very different as a standalone post versus as a reply to "How's your space program going?"
Layer 4: Confidence Scoring
Each potential signal gets a confidence score:
| Confidence Level | Score Range | Action |
|---|---|---|
| High | 80-100% | Alert + Auto-trade eligible |
| Medium | 50-79% | Alert (user decides) |
| Low | 20-49% | Logged, no alert |
| Noise | 0-19% | Filtered out |
Factors Affecting Confidence: - Clarity of language (+) - Specific asset mentioned (+) - Verifiable information (+) - Account credibility (+) - Ambiguity (-) - Humor indicators (-) - Hypothetical language (-)
Specific Signal Types
Exchange Listing Detection
High Confidence Signals:
"@binance: We're excited to announce the listing of $TOKEN"
- Official account ✓
- Clear announcement language ✓
- Specific token ✓
Confidence: 95%+
Lower Confidence:
"Heard $TOKEN might be listing on Binance soon"
- Unofficial source
- Uncertainty ("might," "heard")
- No verifiable info
Confidence: 30-40%
Trade Call Detection
High Confidence:
"Just went long BTC at $50,000, target $55,000, stop at $48,000"
- Clear action
- Specific entry
- Defined targets
Confidence: 90%+
Lower Confidence:
"BTC looking bullish here"
- Vague
- No specific action
- No levels
Confidence: 40-50%
News Event Detection
High Confidence:
"BREAKING: SEC approves Bitcoin ETF"
- News format
- Specific event
- Verifiable
Confidence: 85%+
Lower Confidence:
"ETF approval coming any day now"
- Prediction, not news
- No specific timing
- Speculative
Confidence: 25-35%
Handling Edge Cases
Sarcasm and Humor
Our models detect common sarcasm patterns:
"Oh great, another crypto crash, exactly what I wanted"
Sentiment keywords: great, wanted (+)
Sarcasm indicators: "Oh great," negative context
Classification: Negative sentiment (correctly inverted)
Ambiguous Mentions
When multiple interpretations are possible:
"XRP is a security... a security for my retirement"
Dual meaning detected
Both interpretations flagged
Human review may be triggered
Thread Analysis
For replies and quote tweets:
Original: "What do you think about BTC?"
Reply: "All in."
Without context: Ambiguous
With context: Clear bullish response to BTC question
Classification: Bullish signal on BTC
Our AI models improve over time. User feedback on signal quality helps refine detection accuracy for edge cases.
User Customization
Adjusting Sensitivity
Users can tune the filtering:
Higher Sensitivity: - More signals, more potential noise - Best for: Monitoring, research, manual review
Lower Sensitivity: - Fewer signals, higher quality - Best for: Automated trading, alert fatigue reduction
Custom Keywords
Add your own:
Include Keywords: - Specific tokens you trade - Terms relevant to your strategy - Names of key people/projects
Exclude Keywords: - Tokens you don't trade - Topics that create false positives - Personal terms (humor, memes)
Account-Level Settings
Per-account configuration:
@InfluencerA:
- High priority
- Auto-trade eligible
- Strict filtering (only clear signals)
@NewsAccount:
- Medium priority
- Alert only
- Broad filtering (catch more news)
Performance Metrics
How We Measure Quality
Precision: Of signals flagged, what % were actually relevant? - Target: >90% - Too many false positives = alert fatigue
Recall: Of actual signals, what % did we catch? - Target: >95% - Missing signals = missed opportunities
Speed: How fast from post to classified signal? - Target: <2 seconds - Speed affects tradability
Transparency
In your TradeFollow dashboard: - See why each signal was classified - View confidence scores - Review filtered-out content - Report misclassifications
Conclusion
Effective AI filtering transforms the social media firehose into actionable intelligence:
- Keyword detection catches potential signals
- NLP understanding interprets intent and meaning
- Context analysis provides crucial background
- Confidence scoring prioritizes quality
The result: You receive alerts when there's actually something to act on, not every time someone mentions Bitcoin.
TradeFollow's AI is designed to be your first filter—catching the signals humans would catch, but faster and more consistently. Combined with your judgment and trading rules, it creates a powerful system for capitalizing on social signals.