AI Trading Platform

1 Offline Model Training

Neural networks (LSTM and Transformer) are trained offline on historical data using CVR/MFE soft labels instead of simple win/loss categories. This means the models learn from all directionally correct trades, not just perfect exits.

• Models are frozen during validation and paper execution
• No live retraining — changes require evidence from paper trading
• Soft labels assign graduated scores based on how far price moved in your favor

2 Signal Generation & Aggregation

When analyzing a symbol, the system calculates technical indicators and feeds them to both neural networks. The outputs are combined with learned indicator weights to produce a final signal with confidence score.

• Indicator values are features, not trading rules
• Neural networks provide probability estimates
• Signal Aggregator combines all inputs with calibrated weights

3 Signal-Family Routing

Different market conditions need different exit strategies. The system classifies signals by regime characteristics:

• Breakout signals (high ADX, volume spike) → wider trailing, run to maximize gains
• Mean-reversion signals (ranging market) → tighter trailing, take profits quickly
• Classification uses ADX level, volume ratio, and Bollinger Band position

Signal-family routing is currently under validation during the paper trading phase.

4 Two-Stage Exit Strategy

Research showed trailing stops were cutting winners short by 0.81R per winning trade. The solution splits each position into two parts:

• Core (50%): Exit at 1R profit to lock in gains
• Runner (50%): Uses 1.5R trailing distance to capture larger moves
• This improved expectancy from near-zero to statistically meaningful levels