Risk Management

Take Profit (+5%)

Every position has a hard +5% take profit that fires automatically. When triggered, the actual price move is captured — if the market gapped past 5%, you get the full move. Profits get locked in.

Stop Loss (-3%, Clamped)

A hard -3% stop loss fires automatically on every position. The loss is clamped to exactly -3% regardless of how far the market moved against you — simulating a real stop-loss order. No slippage beyond the stop.

Overtrading Prevention

Two layers: a 5-candle (5-hour) cooldown between trades, and a hard cap of 8 trades per episode. This forces high-quality entries only and keeps fee drag minimal. The agent learned to be selective, not active.

Fee-Aware Decision Making

Every trade is charged a 0.1% fee on entry and exit, applied to the 95% position size. The model is trained with these fees baked in — every P&L number is after costs. No pre-fee inflation.

Max Hold Duration (100 candles)

Positions held longer than 100 candles (100 hours) get force-closed with a penalty. This prevents the model from sitting indefinitely in a losing position hoping for a turnaround.

Data Integrity

The scaler is fitted on training data only — no future price information leaks into the model. Data is split chronologically: 70% train, 15% validation, 15% test. The LSTM validates with early stopping.

Reward Hacking Detection

If the agent produces 10 consecutive episodes with zero trades, the RL policy is automatically reset. If validation shows 3 consecutive zero-trade runs or 2 consecutive weak results, the policy resets. This prevents the model from gaming the reward signal.