Files
gogo2/ANNOTATE/REALTIME_INFERENCE_GUIDE.md
2025-10-25 16:35:08 +03:00

286 lines
6.2 KiB
Markdown

# Real-Time Inference Guide
## 🎯 Overview
Real-time inference mode runs your trained model on **live streaming data** from the DataProvider, updating charts every second and displaying model predictions in real-time.
---
## Starting Real-Time Inference
### Step 1: Select Model
Choose the model you want to run from the dropdown in the training panel.
### Step 2: Click "Start Live Inference"
- Button turns red: "Stop Inference"
- Live mode banner appears at top
- Charts begin updating every second
- Model predictions displayed
### Visual Indicators
- **🔴 LIVE banner** at top of page
- **Green status box** in training panel
- **Update counter** showing number of updates
- **Signal markers** on charts (🔵 BUY, 🔴 SELL)
---
## 📊 What Updates in Real-Time
### Charts (Every 1 Second)
- **All 4 timeframes** update with latest data
- **Candlesticks** show new price action
- **Volume bars** update with new volume
- **Smooth updates** without page refresh
### Model Signals
- **Latest prediction** displayed (BUY/SELL/HOLD)
- **Confidence level** shown as percentage
- **Signal markers** added to charts
- **Last 10 signals** kept visible
### Data Source
- Uses **DataProvider's cached data**
- Same data as main trading system
- Updates from exchange feeds
- 1-second resolution
---
## 🎨 Visual Elements
### Live Mode Banner
```
🔴 LIVE | Real-Time Inference Active
Charts updating with live data every second
[X updates]
```
### Signal Markers on Charts
- **🔵 BUY** - Green marker with arrow
- **🔴 SELL** - Red marker with arrow
- **Timestamp** - When signal was generated
- **Price** - Price at signal time
### Training Panel Status
```
🔴 LIVE
Signal: BUY
Confidence: 75.3%
Charts updating every 1s
```
---
## 🛑 Stopping Real-Time Inference
### Click "Stop Inference"
- Live mode banner disappears
- Charts stop updating
- Signal markers remain visible
- Can review final signals
### What Happens
- Inference loop terminates
- Chart updates stop
- Last 100 signals saved
- Model remains loaded
---
## 📈 Monitoring Performance
### Watch For
- **Signal frequency** - How often model signals
- **Confidence levels** - Higher is better (>70%)
- **Signal accuracy** - Do signals make sense?
- **False positives** - Signals that shouldn't happen
### Good Signs
- Signals at key levels (support/resistance)
- High confidence (>70%)
- Signals match your analysis
- Few false positives
### Warning Signs
- Too many signals (every second)
- Low confidence (<50%)
- Random signals
- Signals don't match patterns
---
## 🔧 Technical Details
### Update Frequency
- **Charts**: 1 second
- **Signals**: 1 second
- **Model inference**: 1 second
### Data Flow
```
DataProvider (Live Data)
Latest Market State (4 timeframes)
Model Inference
Prediction (Action + Confidence)
Update Charts + Display Signal
```
### Performance
- **Latency**: ~100-200ms per update
- **CPU Usage**: Moderate (model inference)
- **Memory**: Stable (no leaks)
- **Network**: Minimal (uses cached data)
---
## 💡 Tips & Tricks
### Tip 1: Watch Multiple Timeframes
All 4 charts update simultaneously. Watch for:
- Alignment across timeframes
- Divergences between timeframes
- Pattern confirmation
### Tip 2: Monitor Confidence
- **>80%**: Very strong signal
- **70-80%**: Strong signal
- **50-70%**: Moderate signal
- **<50%**: Weak signal (ignore)
### Tip 3: Compare with Annotations
- Do live signals match your annotations?
- Are signals at similar price levels?
- Is timing similar to your trades?
### Tip 4: Test Different Models
- Try CNN vs DQN vs Transformer
- Compare signal quality
- Note which performs best
### Tip 5: Use for Validation
- After training, test with live inference
- Verify model learned correctly
- Check for overfitting
---
## 🐛 Troubleshooting
### Charts Not Updating
**Issue**: Live mode active but charts frozen
**Solutions**:
- Check browser console for errors
- Verify DataProvider has live data
- Refresh page and restart inference
- Check network tab for failed requests
### No Signals Generated
**Issue**: Status shows "HOLD" constantly
**Solutions**:
- Model may need more training
- Check model is loaded correctly
- Verify market conditions (model may correctly hold)
- Try different model
### Signals Too Frequent
**Issue**: Signal every second
**Solutions**:
- Model may be overtrained
- Need more negative examples in training
- Adjust confidence threshold
- Retrain with better annotations
### Performance Issues
**Issue**: Browser slow/laggy
**Solutions**:
- Close other tabs
- Reduce number of visible timeframes
- Stop inference when not needed
- Clear browser cache
---
## 📊 Example Session
### Scenario: Testing CNN After Training
**1. Preparation**
- Trained CNN on 20 breakout annotations
- Model learned breakout patterns
- Ready to test on live data
**2. Start Inference**
- Select "StandardizedCNN"
- Click "Start Live Inference"
- 🔴 LIVE banner appears
- Charts begin updating
**3. Observation (5 minutes)**
- Charts update smoothly
- Model generates 2 BUY signals
- Both at resistance breakouts
- Confidence: 78% and 82%
**4. Validation**
- Signals match training patterns
- Timing is precise
- No false positives
- Model learned correctly
**5. Stop Inference**
- Click "Stop Inference"
- Review signal history
- Model performs well
- Ready for production
---
## 🎯 Best Practices
### Before Starting
- Train model first
- Verify model loaded
- Check DataProvider has data
- Close unnecessary tabs
### During Inference
- Monitor all timeframes
- Note signal quality
- Check confidence levels
- Compare with your analysis
### After Stopping
- Review signal history
- Note performance
- Identify improvements
- Adjust training if needed
---
## Summary
Real-time inference provides:
**Live chart updates** (1/second)
**Model predictions** in real-time
**Signal markers** on charts
**Confidence levels** displayed
**Performance monitoring** built-in
Use it to:
- **Validate training** - Check model learned correctly
- **Test models** - Compare different models
- **Monitor performance** - Track signal quality
- **Debug issues** - Identify problems
**Result**: Confidence that your model works correctly before deploying to production! 🎯