You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

CTHULU TNT v2.0

Trading-Native Transformer - A purpose-built neural architecture for financial time-series prediction.

Model Overview

CTHULU TNT (Trading-Native Transformer) v2.0 is a novel neural architecture designed from first principles specifically for financial market prediction. Unlike general-purpose language models adapted for trading, every component is engineered for market dynamics.

Key Innovations

Innovation	Description
Temporal-Causal Attention	Strict causality with learned temporal decay across micro, trend, and macro timescales
Volatility-Adaptive Temperature	Attention softmax temperature adapts to market volatility in real-time
Mixture of Experts (MoE)	3 specialized experts for trending, ranging, and volatile market regimes
Integrated Risk Prediction	SL/TP as first-class model outputs, not external calculations
Trading-Specific Embeddings	Custom embeddings for price levels, session timing, and market structure

Model Card

Metric	Value
Model ID	`cthulu-tnt-v2.0`
Architecture	Trading-Native Transformer
Parameters	401,248
Model Size (Q8)	~392 KB
Sequence Length	500 bars
Input Features	27 (OHLCV + Indicators + Sentiment + Time)
Inference Latency	<5ms (CPU with AVX2)
Throughput	>200 predictions/sec

Architecture

INPUT [Batch, 500, 27]
    │
    ├── Price Embedding (5 → 32)
    ├── Indicator Embedding (10 → 32)
    ├── Sentiment Embedding (7 → 16)
    └── Time Embedding (5 → 16)
           │
           ▼
    Combined: 96 dimensions (d_model)
           │
    ┌──────┴──────┐
    │  TRANSFORMER │ × 3 layers
    │  BACKBONE    │
    │  • Pre-LN    │
    │  • 4 heads   │
    │  • RoPE      │
    │  • SwiGLU    │
    └──────┬──────┘
           │
    ┌──────┴──────┐
    │  MIXTURE OF │
    │   EXPERTS   │
    │  • Trend    │
    │  • Range    │
    │  • Volatile │
    └──────┬──────┘
           │
    ┌──────┴──────┐
    │   OUTPUT    │
    │   HEADS     │
    ├─────────────┤
    │ Direction   │ → 3-class (long/short/neutral)
    │ Price Delta │ → 3 horizons (1, 5, 20 bars)
    │ Confidence  │ → Bayesian uncertainty
    │ Risk        │ → SL/TP in ATR units
    └─────────────┘

Parameter Breakdown

Component	Parameters
Embeddings	672
Transformer Layers (×3)	360,000
MoE Router	288
Expert FFNs (×3)	39,168
Output Heads	960
Layer Norms	160
TOTAL	401,248

Performance Targets

Metric	Target	Description
Direction Accuracy	>52%	Better than random baseline
Profit Factor	>1.3	Gross profit / gross loss
Sharpe Ratio	>0.5	Risk-adjusted return
Max Drawdown	<15%	Worst peak-to-trough
Calibration Error	<5%	Confidence matches accuracy

Usage

Python Integration

from cthulu.model.arch_2.0 import CthulhuTNT, CthulhuConfig

# Load model
config = CthulhuConfig(seq_len=500)
model = CthulhuTNT(config)
model.load("path/to/model.ctml")

# Prepare input: [batch, seq_len, features]
# Features: OHLCV (5) + Indicators (10) + Sentiment (7) + Time (5) = 27
input_data = prepare_market_data(...)

# Inference
outputs = model.forward(input_data)

# Results
direction = outputs['direction_probs']  # [batch, 3] - long/neutral/short
price_delta_1 = outputs['price_delta_1']  # [batch, 1] - 1-bar ahead
price_delta_5 = outputs['price_delta_5']  # [batch, 1] - 5-bars ahead
price_delta_20 = outputs['price_delta_20']  # [batch, 1] - 20-bars ahead
confidence = outputs['confidence']  # [batch, 1] - prediction confidence
stop_loss = outputs['stop_loss']  # [batch, 1] - suggested SL in ATR
take_profit = outputs['take_profit']  # [batch, 1] - suggested TP in ATR

MetaTrader 5 Integration

from cthulu.model.arch_2.0.mt5_bridge import MT5TNTBridge

# Initialize bridge
bridge = MT5TNTBridge(
    model_path="model.ctml",
    symbol="EURUSD",
    timeframe="H1"
)

# Get trading signal
signal = bridge.get_signal()
# Returns: direction, confidence, stop_loss, take_profit

# Full prediction with all outputs
prediction = bridge.predict()

Input Features

Price Features (5)

Open, High, Low, Close, Volume (normalized)

Technical Indicators (10)

RSI (14-period)
MACD (12, 26, 9)
ADX (14-period)
ATR (14-period)
Bollinger Band %B
Stochastic %K, %D
Volume MA ratio
Price momentum
Volatility ratio

Sentiment Features (7)

Market regime indicator
Trend strength
Volatility state
Session activity
Correlation factor
News sentiment (if available)
Order flow imbalance

Time Features (5)

Hour of day (cyclical)
Day of week (cyclical)
Session indicator (Asian/London/NY)

Training

Data Requirements

Minimum: 5 years of M1/M5 data
Recommended: 10+ years across multiple pairs
Format: OHLCV with timestamps

Training Command

cd model/arch_2.0
python train_tnt.py \
    --epochs 100 \
    --batch-size 64 \
    --learning-rate 0.0001 \
    --data-source mt5 \
    --symbols EURUSD,GBPUSD,USDJPY \
    --dashboard

Training Configuration

TrainingConfig(
    learning_rate=1e-4,
    weight_decay=0.01,
    batch_size=32,
    epochs=100,
    lr_schedule="warmup_cosine",
    warmup_epochs=5,
    early_stopping=True,
    patience=10
)

File Format: CTML

Custom binary format optimized for trading models:

CTML File Structure:
├── Header (64 bytes)
│   ├── Magic: "CTML"
│   ├── Version
│   ├── Section offsets
│   └── CRC64 checksum
├── Metadata (JSON)
│   ├── Architecture config
│   ├── Training info
│   └── Performance metrics
├── Feature Config (JSON)
│   ├── Input normalization
│   └── Output specifications
├── Tensor Index
│   └── Name, shape, dtype, offset
└── Tensor Data
    └── Quantized weights (Q8_0)

Deployment Options

Option	Use Case	Performance
Python FastAPI	Development, prototyping	Good
C/C++ Library	Production, low-latency	Best
WebAssembly	Browser deployment	Good
MT5 DLL	Direct MetaTrader integration	Best

Comparison with Other Models

Model	Params	Latency	Trading-Native	Risk Output
CTHULU TNT v2.0	401K	<5ms	✅ Yes	✅ Yes
TimeGPT	200M	~100ms	❌ No	❌ No
Chronos	710M	~500ms	❌ No	❌ No
PatchTST	1M	~10ms	❌ No	❌ No
Informer	10M	~50ms	❌ No	❌ No

CTHULU Advantages

Purpose-Built: Every component designed for trading, not adapted
Low Latency: Sub-5ms inference enables real-time trading
Compact: 401K params vs millions in generic models
Risk-Aware: Native SL/TP prediction, not post-hoc calculation
Regime-Adaptive: MoE handles different market conditions

Limitations & Risks

⚠️ Trading Financial Instruments Carries Substantial Risk

Model predictions are probabilistic, not certain
Past performance does not guarantee future results
Always use proper risk management
Test thoroughly on demo accounts before live trading
The authors are not responsible for trading losses

Known Limitations

Requires quality data preprocessing
Performance may degrade in extreme market conditions
Not designed for HFT (sub-millisecond) applications
Sentiment features require external data sources

Citation

@software{cthulu_tnt_2026,
  title = {CTHULU TNT: Trading-Native Transformer},
  author = {Shakil, Ali A. and Artifact Research},
  year = {2026},
  version = {2.0.0},
  url = {https://huggingface.co/amuzetnoM/CTHULU}
}

License

AGPL-3.0 - See LICENSE

amuzetnoM
/

CTHULU