Autonomous
trading,
zero human
intervention.

LSTM neural networks generate signals across 72 crypto pairs.
A fine-tuned 8B parameter AI validates every entry.
Adaptive risk management protects capital 24/7.
The engine never sleeps.

72
Trained Models
3min
Signal Cycle
24/7
Autonomous
cortex_trading_engine.py

Real-time engine metrics

Every number computed from live exchange data. No estimates. No projections. Raw verified state.

Signal Cycles Run
40+
since activation
Models Trained
72
retrained daily 03:00 UTC
Top Model Accuracy
76.4%
BNB/EUR cross-validated
Open Positions
3/3
max concurrent
Active Positions
LIVE
Pair Side Entry Stop Loss Confidence Val Accuracy Status
LNK
LINK/EUR
Chainlink
LONG €8.076 €7.672
80.4%
70.8% ● OPEN
ADA
ADA/EUR
Cardano
LONG €0.2393 €0.2270
79.0%
70.4% ● OPEN
LTC
LTC/EUR
Litecoin
LONG €49.09 €46.64
77.1%
67.2% ● OPEN

Signal to execution
in under 6 seconds

Every trade passes through a multi-stage pipeline. LSTM generates the signal. A fine-tuned AI validates it. Risk management gates the entry. The engine executes.

01

Market Data Ingestion

6 months of 1H OHLCV candles across 72 pairs. Downloaded daily, appended, and validated for continuity.

02

Feature Engineering

16 technical features extracted per candle: returns, volatility, RSI, Bollinger position, MACD, volume ratios, ATR, moving averages, candle patterns.

03

LSTM Signal Generation

48-bar sequence lookback, 6-bar forward prediction. 128 hidden units, 2 layers, dropout 0.3. Signals above 65% confidence pass through.

04

AI Validation Gate

caelum-8b — a fine-tuned 8B parameter model running locally on GPU — reviews every signal. Disagreements block entry. No exceptions.

05

Execution & Risk Management

Kraken API execution with pre-set stop losses. Position sizing at 2% of capital. Max 3 concurrent positions. 72h max hold time.

Technical Features (16)

  • Price returns & log returns
  • Rolling volatility (20-bar window)
  • Relative Strength Index (RSI-14)
  • Bollinger Band position (% B)
  • MACD signal line & histogram
  • Volume ratio vs 20-bar average
  • Average True Range (ATR)
  • SMA & EMA cross signals
  • Candle body ratio & shadow analysis

AI Validation Layer

  • Fine-tuned 8B parameter LLM
  • Local GPU inference — 6 second latency
  • Reviews market context beyond pure technicals
  • Veto power over LSTM signals
  • Confidence rating: high / medium / low
  • Blocks overfit or anomalous patterns

Execution Rules

  • Min. confidence threshold: 65%
  • Min. validation accuracy: 60%
  • Min. 24h volume: €50,000
  • Max 3 trades per day
  • Max 3 concurrent positions
  • Sacred capital floor: never breached

72 LSTM models.
Retrained every 24 hours.

Each pair gets its own LSTM neural network. Models train on 6 months of hourly data with 50 epochs and early stopping. Cross-validated accuracy shown below.

Capital preservation
is non-negotiable.

The engine is designed to survive, not just profit. Multiple independent circuit breakers ensure no single trade, no single day, and no single failure can compromise the portfolio.

-5%
Hard Stop Loss
Every position enters with a pre-set stop loss. No discretionary overrides. No "let it ride." The stop is the contract between the engine and the market.
2%
Position Sizing
Each trade risks exactly 2% of total capital. No exceptions. This means the engine can absorb 50 consecutive losses before reaching critical drawdown.
72h
Max Hold Time
No position stays open longer than 72 hours. The engine does not become attached to losing positions. Time is a risk factor — it's managed like any other.
Circuit Breakers
ALWAYS ON
Trigger Threshold Action Recovery
24h Drawdown -€30 Halt all trading Next calendar day
Daily Trade Limit 3 trades Block new entries 24h rolling window
Position Limit 3 concurrent Queue signals On position close
Sacred Floor €250 Emergency halt Manual restart only
AI Veto Disagree Block entry Next signal cycle

Latest generated signals

Signals above the 60% confidence threshold from the most recent engine cycle. Only signals that also pass AI validation proceed to execution.

07:31 UTC
ETH/EUR
61.6%
SHORT
58.4% val
07:31 UTC
XRP/EUR
61.2%
SHORT
64.8% val
07:31 UTC
BTC/EUR
60.2%
SHORT
66.4% val
Signal cycle #40  ·  72 models evaluated  ·  3 above threshold

Built to run forever.

The entire stack runs locally on dedicated hardware. No cloud dependencies for trading execution. GPU-accelerated inference. Sub-second order routing.

Compute LOCAL

GPU NVIDIA RTX 4070 SUPER
ML Framework PyTorch + CUDA
LLM Runtime Ollama (localhost:11434)
Validation Model caelum-8b (fine-tuned)
Inference Latency ~6s

Trading Pipeline 24/7

Exchange Kraken (REST + WebSocket)
Signal Cycle Every 3 minutes
Retrain Frequency Daily @ 03:00 UTC
Watchdog 5-min heartbeat check
Eligible Pairs 10 (€50k+ vol)

Data Pipeline DAILY

Candle Resolution 1 Hour
History Depth 6 months
Feature Count 16 per candle
Training Epochs 50 (early stop @ 8)
Download Time 02:00 UTC daily

Resilience FAILSAFE

Process Monitor tmux + watchdog
Auto-Restart Enabled
Health Check Every 10 minutes
Dead Man's Switch Active
State Backup Every 30 minutes