Back to Research

ENSEMBLE + OPUS 4.5

Part 7: The Final Architecture

Omega Arena • February 2026 • ARCHITECTURE DESIGNED

INPUT APPS

171

METRICS

ML MODELS

LLM BRAIN

Abstract. A multi-layered ensemble system combining four independent analytical approaches— the Omega Signal Engine (171 metrics), ML price prediction models (XGBoost, CatBoost, LightGBM, TFT), regime detection models, and real-time news sentiment analysis—all feeding into Claude Opus 4.5 as the ultimate decision maker. The LLM synthesizes conflicting signals, applies portfolio context, and generates actionable trading decisions.

1. Architecture Overview 2. App 1: Signal Generation Engine 3. App 2: ML Price Prediction 4. App 3: Regime Detection 5. App 4: News Sentiment 6. The LLM Decision Layer 7. Data Flow Summary 8. Implementation Roadmap 9. Key Design Decisions 10. Technical Considerations

1. ARCHITECTURE OVERVIEW

APP 1: SIGNAL ENGINE

171 Metrics, Dynamic Params

APP 2: ML MODELS

XGBoost, CatBoost, LightGBM, TFT

APP 3: REGIME

HMM, Random Forest, LSTM

APP 4: NEWS SENTIMENT

RSS + NewsData.io, GPT-4o

↓ ↓ ↓ ↓

OPUS 4.5 (LLM)

Ultimate Decision Maker

↓

TRADING DECISIONS

Coins, Sizes, Stops, Targets

Each APP operates independently, processing market data through different analytical lenses. APPs 1-3 analyze historical price/volume data, while APP 4 provides forward-looking context from real-time news. All outputs are structured JSON that feeds into the LLM for final synthesis.

2. APP 1: SIGNAL GENERATION ENGINE

The Omega System from Part 3—171 technical and fundamental metrics with dynamic parameter adjustment.

Component	Description
171 Metrics	Technical, quantitative, professional, ML-grade, macro, leading
6 Timeframes	1m, 5m, 15m, 30m, 1h, 1d with weighted aggregation
Dynamic Params	Parameters adjust based on detected market conditions
Output	Composite score (-1 to +1), signal (BULLISH/BEARISH/NEUTRAL), confidence

Output Format

{
  "asset": "BTC",
  "timestamp": "2026-02-03T10:00:00Z",
  "composite_score": 0.72,
  "signal": "BULLISH",
  "confidence": 0.68,
  "individual_signals": {
    "rsi_signal": 0.6,
    "macd_signal": 0.8,
    "volume_signal": 0.7,
    // ... 168 more metrics
  }
}
        

3. APP 2: ML PRICE PREDICTION MODELS

Four machine learning models trained on 10 years of data (2015-2025) for price direction prediction.

Model	Type	Best AUC	Status
XGBoost	Gradient Boosting	0.566	DONE
CatBoost	Gradient Boosting	0.530	DONE
LightGBM	Gradient Boosting	0.520	DONE
TFT	Neural Network	N/A	DONE

Output Format

{
  "asset": "BTC",
  "predictions": {
    "xgboost": { "direction": 1, "probability": 0.58 },
    "catboost": { "direction": 1, "probability": 0.54 },
    "lightgbm": { "direction": 0, "probability": 0.52 }
  },
  "ensemble_vote": {
    "direction": 1,
    "agreement": 0.75,
    "weighted_probability": 0.56
  }
}
        

4. APP 3: REGIME DETECTION MODELS

Specialized models dedicated ONLY to detecting market regime. These do NOT predict price—they classify the market state.

Regime	Characteristics	Typical Duration
BULL	Sustained uptrend, expanding volume, positive sentiment	Months to years
BEAR	Sustained downtrend, capitulation events, negative sentiment	Months to years
SIDEWAYS	Range-bound, decreasing volatility, mixed signals	Weeks to months
TRANSITION	Regime change in progress, high uncertainty	Days to weeks

Regime Detection Models (Part 5)

Model	Configuration	Status
Hidden Markov Model (HMM)	219 numeric features, 2-5 states, BIC selection	READY
Random Forest Classifier	235 features, 100 HPO trials, 5-fold CV	READY
BiLSTM + Attention	90-day sequences, 3 layers, multi-task (daily/weekly/monthly)	READY
Ensemble Voting	Weighted voting from all three models	PENDING

Part 5 Dataset

Metric	Value
Assets	97 crypto assets
Raw Features	203 (235 after encoding)
Total Rows	233,507
Date Range	2014-09-17 → 2026-01-26
Labels	100% hindsight-accurate (UP/DOWN/SAME)
LSTM Sequences	90-day windows, 17.6 GB total

Level 7 Context Features. Part 5 introduced 15 new context-aware features: ATH distance, 52-week range position, YTD/yearly returns, seasonality patterns, and Bitcoin halving cycle. These provide the "bigger picture" context that standard technical indicators miss.

Output Format

{
  "regime_detection": {
    "current_regime": "BULL",
    "confidence": 0.78,
    "probabilities": {
      "bull": 0.78,
      "bear": 0.12,
      "sideways": 0.08,
      "transition": 0.02
    },
    "transition_risk": 0.15
  },
  "model_votes": {
    "hmm": "BULL",
    "random_forest": "BULL",
    "lstm": "BULL"
  }
}
        

5. APP 4: NEWS SENTIMENT ANALYSIS

Real-time news analysis providing forward-looking context that technical indicators cannot capture. Built in Part 6.

Why News? Technical indicators analyze the past. News analyzes the future. Markets move on expectations, not just technicals. A Fed announcement, exchange hack, or regulatory change can invalidate all technical signals.

Component	Description
6 RSS Feeds	CoinDesk, CoinTelegraph, Bitcoin Magazine, Decrypt, The Block, Google News
25 Coins	BTC, ETH, SOL, and 22 diversified alts via NewsData.io
Dual-Mode	Quick (10m, GPT-4o-mini) + Full (6h, GPT-4o)
Source Tiers	T1 (1.0x), T2 (0.8x), T3 (0.5x) credibility weighting
Output	Sentiment (BULLISH/BEARISH/NEUTRAL), confidence, key events, risks

Output Format

{
  "news_sentiment": {
    "market_sentiment": "BEARISH",
    "confidence": 0.72,
    "key_events": [
      "Fed signals hawkish stance",
      "Major exchange reports liquidity issues"
    ],
    "asset_outlook": {
      "BTC": "BEARISH",
      "ETH": "NEUTRAL",
      "ALTS": "BEARISH"
    },
    "risk_factors": ["Regulatory uncertainty", "Macro headwinds"],
    "actionable_signal": "SHORT"
  },
  "freshness": {
    "rss_items": 45,
    "newsdata_items": 120,
    "analysis_time": "2026-02-05T14:35:09Z"
  }
}
        

6. THE LLM DECISION LAYER (OPUS 4.5)

The ultimate decision maker. Claude Opus 4.5 receives outputs from all three APPs and makes final trading decisions.

Why an LLM? Unlike traditional ensemble voting, an LLM can:

• Synthesize conflicting signals with nuance
• Request additional raw data when uncertain
• Explain reasoning in natural language
• Apply portfolio context (budget, positions, risk)
• Adapt to changing conditions via prompt updates

LLM Responsibilities

Step	Action
1. Receive	Outputs from APP 1, APP 2, APP 3, APP 4
2. Analyze	Correlations and contradictions between signals
3. Request	Additional raw data if needed (candles, funding rates, etc.)
4. Consider	Portfolio state: budget, open positions, risk tolerance
5. Output	Actionable decisions: BUY/SELL/HOLD with sizes and stops

Input Context

{
  "app1_signals": { /* Signal Engine output */ },
  "app2_predictions": { /* ML Models output */ },
  "app3_regime": { /* Regime Detection output */ },
  "app4_news": { /* News Sentiment output */ },
  "portfolio_state": {
    "total_budget": 100000,
    "available_cash": 45000,
    "open_positions": [
      { "asset": "BTC", "size": 0.5, "entry": 42000, "pnl": "+8%" },
      { "asset": "ETH", "size": 5.0, "entry": 2200, "pnl": "-3%" }
    ],
    "risk_tolerance": "moderate"
  }
}
        

Decision Output

{
  "decisions": [
    {
      "action": "BUY",
      "asset": "SOL",
      "size_usd": 5000,
      "stop_loss": 95.50,
      "take_profit": 125.00,
      "reasoning": "Strong bullish from APP1 (0.72), XGBoost up (0.58), regime BULL."
    },
    {
      "action": "HOLD",
      "asset": "BTC",
      "reasoning": "Already positioned. Signals mixed. Wait for clarity."
    },
    {
      "action": "REDUCE",
      "asset": "ETH",
      "size_reduction": 0.5,
      "reasoning": "Negative PnL, APP1 weakness, transition risk rising."
    }
  ],
  "data_requests": [
    "Need 4h candles for SOL last 7 days",
    "Show me ETH funding rates"
  ]
}
        

7. DATA FLOW SUMMARY

Stage	Input	Process	Output
Raw Data	Price, Volume, On-chain	Collection & Cleaning	Standardized feeds
APP 1	Raw data	171 metrics, signal aggregation	Composite score + signal
APP 2	Raw data	ML model inference	Direction predictions
APP 3	Raw data	Regime classification	Bull/Bear/Sideways
APP 4	RSS + NewsData.io	News sentiment analysis	Bullish/Bearish/Neutral
LLM	All APP outputs + Portfolio	Synthesis + Reasoning	Trading decisions
Execution	LLM decisions	Order routing	Trades placed

8. IMPLEMENTATION ROADMAP

Phase 1: ML Training (Part 4)

Task	Status
CatBoost trained	DONE
LightGBM trained	DONE
TFT trained	DONE
XGBoost (500 HPO trials, AUC 0.566)	DONE

Phase 2: Regime Detection (Part 5)

Task	Status
Dataset: 233K rows × 203 features × 97 assets	DONE
Hindsight labels (UP/DOWN/SAME, BULL/BEAR/SIDEWAYS)	DONE
Level 7 context features (ATH, 52-week, halving cycle)	DONE
ML-ready datasets (train/val/test splits)	DONE
HMM training script	DONE
Random Forest training script	DONE
BiLSTM+Attention training script (90-day sequences)	DONE
Execute training on RunPod	RUNNING
Ensemble voting system	PENDING

Phase 3: News Sentiment (Part 6)

Task	Status
RSS feed integration (6 sources)	DONE
NewsData.io API integration (25 coins)	DONE
Source credibility tiers (T1/T2/T3)	DONE
LLM prompt engineering (structured JSON output)	DONE
Dual-mode daemon (10m quick + 6h full)	TESTING
Price correlation validation	TESTING

Phase 5: Integration Layer

Task	Status
Standardize JSON output schemas	PENDING
Build inference pipeline	PENDING
Create portfolio state manager	PENDING
Implement raw data API	PENDING

Phase 6: LLM Integration

Task	Status
Design Opus 4.5 prompt templates	PENDING
Implement conversation flow	PENDING
Build decision execution layer	PENDING
Add logging and audit trail	PENDING

Phase 7: Testing & Go-Live

Task	Status
Backtest on 2024-2025 data	PENDING
Paper trading phase	PENDING
Risk management fine-tuning	PENDING
Go live with small allocation	PENDING

9. KEY DESIGN DECISIONS

Why separate Regime Detection (APP 3)?

Regime is a higher-level concept than individual predictions
Allows the LLM to weight APP1/APP2 outputs appropriately
Regime models can be simpler and more interpretable

Why use LLM as final decision maker?

Synthesizes conflicting signals with nuance
Handles edge cases and anomalies
Provides explainable reasoning
Requests additional context when uncertain
Adapts via prompt updates without retraining

Why dynamic signal parameters in APP 1?

Static parameters fail in different market conditions
Regime-aware adjustment improves signal quality
Reduces false signals in trending vs ranging markets

Why keep all 4 ML models?

Different models capture different patterns
Ensemble disagreement is informative
Redundancy improves robustness

Why add News Sentiment (APP 4)?

Technical indicators only see the past
News provides forward-looking context
Markets move on expectations, not just technicals
Can invalidate all technical signals (Fed, hacks, regulation)
Hybrid RSS + NewsData.io achieves $0.35/day vs $200/mo paid APIs

10. TECHNICAL CONSIDERATIONS

Consideration	Approach
Part 4 Data (ML Models)	10 years (2015-2025), ~197k rows × 16k features
Part 5 Data (Regime)	233K rows × 203 features × 97 assets
LSTM Memory	90-day sequences, 17.6 GB (requires A100/L40 GPU)
LLM Cost	Opus 4.5 API calls batched/rate-limited
Latency	Real-time decisions need <5s total inference
Failsafe	If LLM unavailable, fall back to simple ensemble voting

PAPER TRADING FIRST: This architecture will be validated through extensive paper trading before any real capital deployment. The LLM adds a layer of uncertainty that requires careful testing.

Architecture + Data Pipeline Complete. This document defines the full 4-layer ensemble system. Part 1 established the baseline. Part 3 documented the Omega System. Part 4 completed ML training (XGBoost AUC: 0.566). Part 5 built the regime detection pipeline: 233K rows × 203 features × 97 assets. Part 6 added real-time news sentiment analysis via hybrid RSS + NewsData.io. Part 7 integrates all four layers with Claude Opus 4.5 as the final decision maker.

ENSEMBLE + OPUS 4.5

TABLE OF CONTENTS

1. ARCHITECTURE OVERVIEW

2. APP 1: SIGNAL GENERATION ENGINE

Output Format

3. APP 2: ML PRICE PREDICTION MODELS

Output Format

4. APP 3: REGIME DETECTION MODELS

Regime Detection Models (Part 5)

Part 5 Dataset

Output Format

5. APP 4: NEWS SENTIMENT ANALYSIS

Output Format

6. THE LLM DECISION LAYER (OPUS 4.5)

LLM Responsibilities

Input Context

Decision Output

7. DATA FLOW SUMMARY

8. IMPLEMENTATION ROADMAP

Phase 1: ML Training (Part 4)

Phase 2: Regime Detection (Part 5)

Phase 3: News Sentiment (Part 6)

Phase 5: Integration Layer

Phase 6: LLM Integration

Phase 7: Testing & Go-Live

9. KEY DESIGN DECISIONS

Why separate Regime Detection (APP 3)?

Why use LLM as final decision maker?

Why dynamic signal parameters in APP 1?

Why keep all 4 ML models?

Why add News Sentiment (APP 4)?

10. TECHNICAL CONSIDERATIONS