AI Trading Bridge for NinjaTrader 8

Phase 3 LLM-Enhanced Model Integration

This Python program connects NinjaTrader 8 to your trained Stable-Baselines3 Phase 3 model with LLM-enhanced features for automated trading on E-mini Nasdaq-100 (NQ).

Architecture

[NinjaTrader 8] <--TCP:8888--> [Python Bridge] <---> [Phase 3 Model + Adapter]
   AIBridgeV2.cs                 ai_trading_bridge.py   NQ_ts-0040960...zip

Data Flow (Phase 3):

NinjaTrader sends real-time bar data (OHLCV + indicators)
Python builds 261D observation:
- 220D market features (20 bars × 11 indicators)
- 8D position state (includes validity flags)
- 33D LLM features (extended market context)
VecNormalize applies normalization (trained on 40,960 timesteps)
Adapter layer reduces 261D → 228D
Policy network predicts action (0-5)
Python sends action to NinjaTrader
NinjaTrader executes the action

Files Created

models/
├── ai_trading_bridge.py        # Main program
├── config.py                   # Configuration settings (Phase 3)
├── tcp_client.py               # TCP communication
├── observation_builder.py      # Observation construction (261D)
├── model_manager.py            # Model loading & inference (with adapter)
├── llm_features.py             # LLM feature builder (33D features)
├── trade_logger.py             # Performance tracking
├── requirements.txt            # Dependencies
├── README.md                   # This file
│
├── model/                                                          # Phase 3 model directory
│   ├── NQ_ts-0040960_evt-best_val-+216.697_sharpe-+0.48_seed-42.zip           # ✅ Trained model
│   ├── NQ_ts-0040960_evt-best_val-+216.697_sharpe-+0.48_seed-42_vecnormalize.pkl  # ✅ VecNormalize
│   └── NQ_ts-0040960_evt-best_val-+216.697_sharpe-+0.48_seed-42_metadata.json     # Metadata
│
└── logs/                       # Auto-created
    ├── trading_log.txt         # Session logs
    ├── actions.csv             # All actions taken
    └── performance.txt         # Statistics

Installation

1. Install Python Dependencies

cd "/mnt/c/Users/javlo/Documents/NinjaTrader 8/bin/Custom/models"
pip install -r requirements.txt

Dependencies:

stable-baselines3>=2.0.0 - RL framework
sb3-contrib>=2.0.0 - Advanced algorithms (MaskablePPO)
numpy>=1.21.0 - Numerical computing
gym>=0.21.0 - Environment interface

2. Prepare Required Files (Phase 3)

✅ Model File (Already Present)

File: model/NQ_ts-0040960_evt-best_val-+216.697_sharpe-+0.48_seed-42.zip
Size: 2.5 MB
Type: MaskablePPO with adapter layer (261D → 228D)
Trained: 40,960 timesteps
Performance: +$216.697 validation, Sharpe 0.48

✅ VecNormalize Stats File (FOUND!)

File: model/NQ_ts-0040960_evt-best_val-+216.697_sharpe-+0.48_seed-42_vecnormalize.pkl
Size: 28.6 KB
Shape: (261,) - Includes LLM features
Trained: 40,960 timesteps (Phase 3)

These files are CRITICAL for model operation and are already present!

3. Validate Configuration

Test that everything is configured correctly:

python config.py

Expected output (Phase 3):

======================================================================
 AI Trading Bridge - Configuration Validation
======================================================================
✅ Configuration valid!

📁 Model path: .../model/NQ_ts-0040960_evt-best_val-+216.697_sharpe-+0.48_seed-42.zip
📁 VecNormalize path: .../model/NQ_ts-0040960..._vecnormalize.pkl
📁 Log directory: .../logs

🔢 Observation shape: (261,)
   - Market features: 220 (20 bars × 11 features)
   - Position features: 8
   - LLM features: 33
   - Base (Phase 2): 228, Extended (Phase 3): 261

🌐 TCP: localhost:8888
📊 Model type: MaskablePPO
======================================================================

If you see errors:

Missing model file → Ensure Phase 3 model exists in model/ subdirectory
Missing VecNormalize file → This is CRITICAL (must match model)

Testing

Test 1: TCP Connection

Test TCP connection to NinjaTrader:

python tcp_client.py

Before running:

Start NinjaTrader 8
Load NQ 1-minute chart
Add AIBridgeV2 strategy to chart
Enable strategy (green arrow)

Expected output:

🔌 Connecting to NinjaTrader @ localhost:8888...
✅ Connected to NinjaTrader successfully!

Waiting for messages (Ctrl+C to stop)...

📨 Received: bar
   Close: 21450.25
   ATR: 45.5

📤 Sending test action: HOLD (0)

If connection fails:

Check NinjaTrader Output Window for "TCP Server started"
Ensure AIBridgeV2 strategy is enabled
Check no firewall blocking port 8888

Test 2: Observation Builder

Test observation construction:

python observation_builder.py

Expected output (Phase 3):

1. Adding first bar...
   Window size: 1/20
⚠️  LLM feature calculation failed (using base obs only): single positional indexer is out-of-bounds
✅ Observation shape: (261,)
   First 11 features (bar 1): [21450.25  1234.  21448. ...]
   Position features (8): [0. 1. 0. 0. 0. 1. 0. 0.]
   LLM features (33): shape=(33,), sample=[0. 0. 0. 0. 0.]

2. Adding 19 more bars...
   Window size: 20/20
✅ Full observation shape: (261,)
   Market features (220): min=0.00, max=21470.00
   Position features (8): [0. 1. 0. 0. 0. 1. 0. 0.]
   LLM features (33): min=0.00, max=0.00

Note: LLM feature warnings are normal during warmup (needs 60+ bars for multi-timeframe calculations).

Test 3: Model Loading

Test model loading and predictions:

python model_manager.py

Expected output (Phase 3):

======================================================================
 Loading Phase 3 AI Model
======================================================================

📁 Loading model: .../model/NQ_ts-0040960...zip
   ⚠️  MaskablePPO detected - trying standard loader first
   ⚠️  Standard loading failed (expected for Phase 3)
   Trying custom loader (handles adapter architecture)...
   🔧 Using custom loader for pickle compatibility...
   Detecting network architecture from weights...
   Architecture: pi=[512, 256, 128], vf=[512, 256, 128]
   📐 Adapter detected: torch.Size([228, 261])
   📐 Phase 3 model - extracting adapter (261D → 228D)
   ✅ Policy weights loaded (228D input)
   ✅ Adapter extracted: torch.Size([228, 261])
✅ Adapter layer created in ModelManager
✅ Model loaded with custom loader: MaskablePPO

📁 Loading VecNormalize: .../model/NQ_ts-0040960..._vecnormalize.pkl
✅ VecNormalize loaded
   Running mean shape: (261,)
   Running var shape: (261,)

✅ Model ready for predictions!

2. Testing prediction with dummy observation...
   Dummy observation shape: (261,)
✅ Prediction successful!
   Action: 4 (ENABLE_TRAIL)

If model loading fails:

Check Phase 3 model file exists in model/ subdirectory
Check VecNormalize file matches model (261D)
Adapter architecture is required for Phase 3 (automatic)

Test 4: Trade Logger

Test logging functionality:

python trade_logger.py

Expected output:

1. Logging sample bars and actions...
[2025-11-07T10:30:00.000Z] 📊 Bar #1 | Close=21450.00 | Vol=1200 | ATR=45.00
[2025-11-07T10:30:00.000Z] 🎯 Action 1: BUY

... (more bars) ...

─────────────────────────────────────────────────────────────────
 Performance Summary (Live)
─────────────────────────────────────────────────────────────────
 Runtime: 0s
 Bars Processed: 15
 Actions Sent: 15

 Action Distribution:
   ✅ HOLD            (0):   12 ( 80.0%) [Expected: 86.5%]
   ✅ BUY             (1):    1 (  6.7%) [Expected: 5.0%]
   ...

Running the Live System

Prerequisites Checklist

Before running live trading:

✅ Python dependencies installed
✅ NQ_o1.zip model file present
✅ phase2_position_mgmt_final_vecnorm.pkl file present ✅
✅ All tests passed (tcp, observation, model, logger)
✅ NinjaTrader 8 running
✅ NQ 1-minute chart loaded
✅ AIBridgeV2 strategy compiled (F5)
✅ AIBridgeV2 strategy enabled on chart
✅ Using Sim101 account (NOT live account!)

Step 1: Start NinjaTrader

Open NinjaTrader 8
Connect to data feed
Load chart: Instruments → NQ 09-25 → 1 Minute
Add strategy: Strategies → AIBridgeV2
Enable strategy (green arrow at top)

Verify in Output Window:

═══════════════════════════════════════════════════════════════
  AIBridgeV2 Starting... (Phase 2 Model Support)
  Integer Actions: 0-5 | ATR-based SL/TP | Trailing Stops
═══════════════════════════════════════════════════════════════
✓ TCP Server started - Waiting for Python client connection...

Step 2: Run Python Bridge

Open a terminal and run:

cd "/mnt/c/Users/javlo/Documents/NinjaTrader 8/bin/Custom/models"
python ai_trading_bridge.py

Expected startup sequence:

======================================================================
 AI Trading Bridge - Phase 2 Model
 Initializing...
======================================================================

1. Validating configuration...
✅ Configuration valid

2. Initializing components...
✅ Observation builder ready
✅ Model manager ready
✅ Logger ready
✅ TCP client ready

3. Loading AI model...

======================================================================
 Loading Phase 2 AI Model
======================================================================
✅ Model loaded: MaskablePPO
✅ VecNormalize loaded

4. Connecting to NinjaTrader...
🔌 Connecting to NinjaTrader @ localhost:8888...
✅ Connected to NinjaTrader successfully!

======================================================================
✅ All systems ready!
======================================================================

======================================================================
 AI Trading Bridge - RUNNING
 Press Ctrl+C to stop
======================================================================

📊 NinjaTrader Configuration:
   Symbol: NQ 09-25
   Tick Size: 0.25
   Max Position: 10

Step 3: Monitor Live Trading

NinjaTrader Output:

[2025-11-07 10:30:00] 📊 Bar data sent: Close=$21450.25, ...
✅ Action sent to NinjaTrader
[2025-11-07 10:30:00] 📨 Received Phase 2 action: 0
[2025-11-07 10:30:00] Action 0: HOLD - No action taken

Python Console Output:

[2025-11-07T10:30:00.000Z] 📊 Bar #1 | Close=21450.25 | Vol=1234 | ATR=45.5
[2025-11-07T10:30:00.000Z] 🎯 Action 0: HOLD
📤 Sending: HOLD (action=0)

[2025-11-07T10:31:00.000Z] 📊 Bar #2 | Close=21452.75 | Vol=1456 | ATR=45.2
   ⏳ Building window: 2/20 bars

[2025-11-07T10:49:00.000Z] 📊 Bar #20 | Close=21478.50 | Vol=1889 | ATR=46.1
[2025-11-07T10:49:00.000Z] 🎯 Action 1: BUY
📤 Sending: BUY (action=1)

When a trade executes:

[2025-11-07T10:49:00.000Z] 🎯 Action 1: BUY
   ✅ Order filled: AI_LONG @ 21478.50 x 1
   📍 Position: LONG @ 21478.50

[2025-11-07T10:52:00.000Z] 🎯 Action 3: MOVE_TO_BE
   ✅ MOVE_TO_BE executed | New SL: 21478.75 | PnL: $600.00 | Count: 1

[2025-11-07T10:55:00.000Z] 🎯 Action 4: ENABLE_TRAIL
   ✅ ENABLE_TRAIL activated | SL: 21533.00 | Profit: $1,390.00 (1.05R)

[2025-11-07T10:56:00.000Z] 📊 Bar #27 | Close=21548.25
   ↑ Trail SL updated: 21548.25 | Profit: $1,695.00

Step 4: Monitor Performance

Every 10 bars, statistics are printed:

─────────────────────────────────────────────────────────────────
 Performance Summary (Live)
─────────────────────────────────────────────────────────────────
 Runtime: 23m 45s
 Bars Processed: 142
 Actions Sent: 142

 Action Distribution:
   ✅ HOLD            (0):  122 ( 85.9%) [Expected: 86.5%]
   ✅ BUY             (1):    7 (  4.9%) [Expected: 5.0%]
   ✅ SELL            (2):    6 (  4.2%) [Expected: 5.0%]
   ✅ MOVE_TO_BE      (3):    4 (  2.8%) [Expected: 2.0%]
   ✅ ENABLE_TRAIL    (4):    2 (  1.4%) [Expected: 1.0%]
   ✅ DISABLE_TRAIL   (5):    1 (  0.7%) [Expected: 0.5%]

 Position Management Usage: 4.9% [Target: ~13.5%]
─────────────────────────────────────────────────────────────────

Indicators:

✅ Green checkmark: Distribution within 5% of expected
⚠️ Yellow warning: Distribution 5-10% off expected
❌ Red X: Distribution >10% off expected

Step 5: Stop Trading

Graceful Shutdown:

Press Ctrl+C in Python terminal
Python will:
- Print final statistics
- Disconnect from NinjaTrader cleanly
- Close log files
- Exit

Final Output:

⚠️  Shutdown signal received...

======================================================================
 Final Session Statistics
======================================================================
 Session Duration: 1h 23m 45s
 Total Bars: 142
 Total Actions: 142

 Action Distribution:
   HOLD            (0):  122 ( 85.9%)
   BUY             (1):    7 (  4.9%)
   SELL            (2):    6 (  4.2%)
   MOVE_TO_BE      (3):    4 (  2.8%)
   ENABLE_TRAIL    (4):    2 (  1.4%)
   DISABLE_TRAIL   (5):    1 (  0.7%)

 Position Management Usage: 4.9%
 Expected PM Usage: 13.5%
 ⚠️  PM usage outside expected range

======================================================================
✅ Shutdown complete
======================================================================

Log Files

All logs are saved in models/logs/:

`trading_log.txt`

Complete session log with timestamps:

======================================================================
 AI Trading Bridge Session Started
 Time: 2025-11-07 10:30:00
======================================================================

2025-11-07T10:30:00 | Action 0: HOLD
2025-11-07T10:49:00 | Action 1: BUY
2025-11-07T10:52:00 | Action 3: MOVE_TO_BE
...

`actions.csv`

Structured action log (importable to Excel):

timestamp,bar_index,action,action_name,close_price,atr,position
2025-11-07T10:30:00,1,0,HOLD,21450.25,45.5,0
2025-11-07T10:49:00,20,1,BUY,21478.50,46.1,0
2025-11-07T10:52:00,23,3,MOVE_TO_BE,21485.00,45.8,1

`performance.txt`

Session summaries and statistics

Troubleshooting

Problem: "Model file not found"

Solution: Ensure NQ_o1.zip exists in models/ directory

Problem: "VecNormalize file not found"

Solution: This is CRITICAL. You must:

Re-run your training script
VecNormalize stats already present: phase2_position_mgmt_final_vecnorm.pkl
Copy file to models/ directory

Without this file, model predictions will be incorrect!

Problem: "Connection refused"

Solution:

Ensure NinjaTrader is running
Ensure AIBridgeV2 strategy is enabled on chart
Check Output Window for "TCP Server started"
Try reloading the strategy

Problem: "Invalid observation shape"

Solution:

Check that NinjaTrader sends all 8 indicators
Verify timestamp format is ISO 8601
Run python observation_builder.py to test

Problem: "PM usage outside expected range"

Possible causes:

Model still warming up (wait for more data)
Market conditions different from training
Position state not updating correctly from NinjaTrader

Not necessarily an error - monitor over longer period

Advanced Configuration

Change Model Type

If you trained with a different algorithm:

Edit config.py:

MODEL_TYPE = "DQN"  # Options: PPO, MaskablePPO, DQN, A2C, SAC

Supported Algorithms:

"PPO" - Proximal Policy Optimization (stable-baselines3)
"MaskablePPO" - PPO with action masking (sb3-contrib) ← Current model
"DQN" - Deep Q-Network (stable-baselines3)
"A2C" - Advantage Actor-Critic (stable-baselines3)
"SAC" - Soft Actor-Critic (stable-baselines3)

Note: MaskablePPO requires sb3-contrib package

Adjust Logging Verbosity

Edit config.py:

VERBOSE = True           # Print detailed output
SHOW_BAR_DATA = True     # Print every bar
SHOW_OBSERVATIONS = False # Print observation arrays (noisy)
SHOW_PREDICTIONS = True   # Print model predictions
STATS_UPDATE_FREQUENCY = 10  # Print stats every N bars

Change TCP Port

If port 8888 is in use:

Edit config.py:

TCP_PORT = 9999  # Or any available port

Edit AIBridgeV2.cs in NinjaTrader:

ServerPort = 9999  // Match Python port

Recompile strategy (F5)

Performance Expectations

Based on Phase 2 model training:

Metric	Expected Value
Sharpe Ratio	~21.71
Returns	+11%
PM Usage	~13.5% of actions
HOLD Actions	~86.5%
Entry Actions (BUY/SELL)	~10%
PM Actions (3,4,5)	~3.5%

Note: Performance may vary based on:

Market conditions
Data feed quality
Execution slippage
Actual vs simulated commissions

Safety Recommendations

⚠️ IMPORTANT SAFETY RULES

ALWAYS test in Sim101 account first - NEVER use live account initially
Run for minimum 1 week in simulation before considering live
Monitor action distribution - if significantly different from expected, investigate
Check log files daily for errors or unusual behavior
Set up alerts for disconnections or errors
Have a manual override plan to close positions if needed

🛑 STOP TRADING IF:

PM usage consistently >20% or <5%
Action distribution deviates >20% from expected
Model predictions become erratic
TCP connection becomes unstable
Unexpected errors in logs

Support

For issues:

Check this README troubleshooting section
Review log files in models/logs/
Test individual components (tcp_client.py, model_manager.py, etc.)
Check NinjaTrader Output Window for errors

Model-specific questions:

Refer to your model training documentation
Check Phase 2 model evaluation results
Verify VecNormalize stats are from correct training run

Quick Reference

Start Trading

# Terminal 1: Start NinjaTrader, enable AIBridgeV2
# Terminal 2:
cd "/mnt/c/Users/javlo/Documents/NinjaTrader 8/bin/Custom/models"
python ai_trading_bridge.py

Stop Trading

Ctrl+C in Python terminal

View Logs

tail -f logs/trading_log.txt     # Real-time log
cat logs/actions.csv              # Action history

Test Components

python config.py              # Validate configuration
python tcp_client.py          # Test TCP connection
python observation_builder.py # Test observation builder
python model_manager.py       # Test model loading
python trade_logger.py        # Test logger

Version Information

Python Bridge Version: 2.0 (Phase 3)
Compatible with: NinjaTrader 8, AIBridgeV2 strategy
Model: Phase 3 LLM-Enhanced (Stable-Baselines3 + Adapter)
Observation Space: (261,) - 220 market + 8 position + 33 LLM features
Action Space: Discrete(6) - Actions 0-5
Model Architecture: MaskablePPO with 261D → 228D adapter layer
Training: 40,960 timesteps, gymnasium framework

Created: November 7, 2025 Last Updated: November 26, 2025 (Phase 3 Migration)

⚠️ DISCLAIMER: This software is for educational and testing purposes only. Trading involves risk. Always test thoroughly in simulation before live trading. Past performance does not guarantee future results.

Name		Name	Last commit message	Last commit date
Latest commit History 2 Commits
NT8		NT8
model		model
FILE_LIST.txt		FILE_LIST.txt
README.md		README.md
ai_trading_bridge.py		ai_trading_bridge.py
config.py		config.py
llm_features.py		llm_features.py
model_manager.py		model_manager.py
observation_builder.py		observation_builder.py
requirements.txt		requirements.txt
tcp_client.py		tcp_client.py
trade_logger.py		trade_logger.py

Folders and files

Latest commit

History

Repository files navigation

AI Trading Bridge for NinjaTrader 8

Architecture

Files Created

Installation

1. Install Python Dependencies

2. Prepare Required Files (Phase 3)

✅ Model File (Already Present)

✅ VecNormalize Stats File (FOUND!)

3. Validate Configuration

Testing

Test 1: TCP Connection

Test 2: Observation Builder

Test 3: Model Loading

Test 4: Trade Logger

Running the Live System

Prerequisites Checklist

Step 1: Start NinjaTrader

Step 2: Run Python Bridge

Step 3: Monitor Live Trading

Step 4: Monitor Performance

Step 5: Stop Trading

Log Files

trading_log.txt

actions.csv

performance.txt

Troubleshooting

Problem: "Model file not found"

Problem: "VecNormalize file not found"

Problem: "Connection refused"

Problem: "Invalid observation shape"

Problem: "PM usage outside expected range"

Advanced Configuration

Change Model Type

Adjust Logging Verbosity

Change TCP Port

Performance Expectations

Safety Recommendations

⚠️ IMPORTANT SAFETY RULES

🛑 STOP TRADING IF:

Support

Quick Reference

Start Trading

Stop Trading

View Logs

Test Components

Version Information

RL_Agent

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

`trading_log.txt`

`actions.csv`

`performance.txt`