NEWS SENTIMENT ANALYSIS
Part 6: Real-Time Context Awareness for AI Trading
TABLE OF CONTENTS
1. THE PROBLEM: BLIND TO THE FUTURE
Parts 3-5 created a powerful analytical engine with 171+ metrics. But all of them share one limitation: they only see the past.
- RSI sees the last 14 days
- MACD sees the last 26 days
- Even LSTM sees the last 90 days
None of these know that the Fed is announcing rates tomorrow, that a major exchange was just hacked, or that a country just banned crypto. News moves markets before technicals can react.
2. FAILED APPROACHES: THE $200/MONTH WALL
Several professional crypto news APIs were evaluated. The results were discouraging.
CryptoPanic API
| Tier | Price | Delay | Requests | Features |
|---|---|---|---|---|
| Free (Dev) | $0 | 24h delay | 100/month | Title only |
| Growth | $199/mo | Real-time | 3,000/month | Full content, sentiment |
| Enterprise | $899/mo | Real-time | Custom | Push API, search |
NewsData.io Crypto API
| Tier | Price | Credits/Day | Delay | Features |
|---|---|---|---|---|
| Free | $0 | 200 | 12h delay | Title, description, source priority |
| Basic | $199/mo | 20,000 | Real-time | Full content, AI summary |
| Professional | $349/mo | 50,000 | Real-time | Sentiment analysis |
source_priority — a numerical
ranking of source credibility based on traffic and authenticity. This is valuable metadata
even with the 12h delay.
Other APIs Evaluated
| API | Status | Issue |
|---|---|---|
| NewsAPI.org | REJECTED | No crypto-specific endpoint |
| CoinGecko | REJECTED | No news API in free tier |
| Messari | REJECTED | Enterprise pricing only |
3. THE SOLUTION: HYBRID FREE ARCHITECTURE
The realization came that two free sources could be combined to achieve professional-grade coverage:
RSS Sources (Tier 1 & 2)
| Source | Tier | Type | Credibility |
|---|---|---|---|
| Google News (Crypto) | T1 | Aggregator | Aggregates major outlets |
| CoinDesk | T1 | Original | Industry standard |
| CoinTelegraph | T2 | Original | Crypto-native |
| Bitcoin Magazine | T2 | Original | BTC-focused |
| Decrypt | T2 | Original | Web3-focused |
| The Block | T1 | Original | Institutional focus |
NewsData.io Free Tier (Enrichment)
- 200 credits/day — Each credit returns up to 10 articles
- 12h delay — Not real-time, but provides detailed descriptions
- Source priority scores — Numerical credibility ranking
- 25 coins tracked — BTC, ETH, SOL, and 22 diversified alts
Source Credibility Tiers
| Tier | Impact Weight | Sources | Rationale |
|---|---|---|---|
| T1 | 1.0x | Reuters, Bloomberg, CoinDesk, The Block | Institutional-grade, fact-checked |
| T2 | 0.8x | CoinTelegraph, Decrypt, Bitcoin Magazine | Crypto-native, generally reliable |
| T3 | 0.5x | BeInCrypto, AMBCrypto, aggregators | Higher speculation, verify claims |
4. DUAL-MODE MONITORING SYSTEM
Two complementary analysis modes run continuously:
Quick Mode (Every 10 Minutes)
| Parameter | Value |
|---|---|
| Data source | RSS feeds only (T1/T2) |
| LLM | GPT-4o-mini |
| Cost per call | ~$0.001 |
| Daily cost | ~$0.15 |
| Purpose | Catch breaking news quickly |
Full Mode (Every 6 Hours)
| Parameter | Value |
|---|---|
| Data source | RSS + NewsData.io (all tiers) |
| LLM | GPT-4o |
| Cost per call | ~$0.05 |
| Daily cost | ~$0.20 |
| Purpose | Deep analysis with enriched context |
5. LLM ANALYSIS PIPELINE
Raw headlines are not useful. They must be synthesized into actionable intelligence. A carefully engineered prompt transforms news into structured signals.
LLM Role Definition
- Skeptical: Assume manipulation until proven otherwise
- Probabilistic: Express confidence levels, not certainties
- Contrarian: Consider opposite interpretations
- Time-aware: Pay attention to publication timestamps
Structured Output
| Field | Type | Description |
|---|---|---|
| market_sentiment | BULLISH/BEARISH/NEUTRAL | Overall market direction |
| confidence | 0-100% | How strongly the LLM believes this |
| key_events | Array | Most impactful news items |
| asset_outlook | Object | Per-asset sentiment (BTC, ETH, ALTS) |
| risk_factors | Array | What could go wrong |
| contrarian_view | String | Devil's advocate perspective |
| actionable_signal | LONG/SHORT/HOLD | Trading recommendation |
Prompt Engineering Principles
- Timestamp awareness: News items include publication times so LLM understands chronology
- Freshness indicators: RSS marked as real-time, NewsData marked as 12h delayed
- Source credibility: Tier information passed to weight importance
- Structured JSON output: Enforced via system prompt for programmatic parsing
6. CURRENT TESTING
A live daemon is running, collecting and analyzing news in real-time. Early results are being validated against actual price movements.
Validation Methodology
- Record BTC price at time of each analysis
- Record predicted direction (BULLISH/BEARISH)
- Check actual price movement in subsequent hours
- Calculate prediction accuracy over time
Testing Status
| Component | Status | Notes |
|---|---|---|
| RSS collection | WORKING | 6 feeds, real-time |
| NewsData.io integration | WORKING | 25 coins, 200 credits/day |
| Quick analysis (10m) | TESTING | GPT-4o-mini |
| Full analysis (6h) | TESTING | GPT-4o |
| Price correlation | TESTING | Manual validation ongoing |
7. COINS TRACKED
NewsData.io free tier limits daily calls. A diversified selection of 25 coins was chosen to maximize coverage without exceeding limits.
| Category | Coins |
|---|---|
| Major | BTC, ETH |
| Layer 1 | SOL, ADA, AVAX, DOT, NEAR, ATOM, SUI, APT, SEI |
| Layer 2 | MATIC, ARB, OP |
| DeFi | UNI, AAVE, LINK, MKR |
| Exchange | BNB |
| Meme (sentiment) | DOGE, SHIB |
| Other | XRP, LTC, TRX |
8. INTEGRATION WITH ENSEMBLE
News sentiment becomes the fourth layer in the Omega ensemble architecture (Part 7).
| Layer | Source | Signal Type |
|---|---|---|
| 1 | Signal Engine (171 metrics) | Technical/Quantitative |
| 2 | ML Models (XGBoost, etc.) | Pattern Recognition |
| 3 | Regime Detection (HMM/RF/LSTM) | Market State |
| 4 | News Sentiment | Forward-Looking Context |
9. LIMITATIONS & FUTURE WORK
Current Limitations
- Headlines only: RSS provides titles, not full articles
- 12h delay: NewsData.io free tier is not real-time
- English only: Non-English news not captured
- No historical training: LLM analysis is zero-shot, not trained on past news-price correlations
Future Improvements
- Build historical dataset of news → price movements
- Train a fine-tuned model for crypto-specific sentiment
- Add social media signals (Twitter/X, Reddit)
- Implement "priced in" detection by correlating news timing with price charts
10. CURRENT STATUS
| Component | Status | Notes |
|---|---|---|
| RSS Pipeline | DONE | 6 sources, unlimited calls |
| NewsData.io Integration | DONE | 25 coins, free tier |
| Source Credibility Tiers | DONE | T1/T2/T3 classification |
| LLM Prompt Engineering | DONE | Structured JSON output |
| Dual-Mode Daemon | TESTING | 10m quick + 6h full |
| Price Correlation Analysis | TESTING | Manual validation |
| Ensemble Integration | PENDING | After testing complete |
Infrastructure complete. Live testing daemon running. Validating predictions against real market movements.
© 2026 Omega Arena