AMERTUME
xLSTM + PPO Reinforcement Learning XAUUSD M15
My backtest ran on fixed 0.01 micro-lots the smallest trade size possible. It proved the model works: 6.94 Sharpe, 40.5% win rate, 0.98% max drawdown over 294 trades on unseen 2024–2025 data.
But 0.01 lots on a $10K account turned $507 profit in 1.87 years. That's 2.7% annually worse than a savings account. To actually make money, I need dynamic position sizing. And dynamic sizing means real risk which needs real protection.
BACKTEST → LIVE: WHAT CHANGED
Instead of fixed 0.01, I risk 1% of balance per trade. Lot size adapts to ATR wide volatility means smaller lot, tighter means bigger. Dollar risk stays constant regardless of market conditions.
lot = (balance × 1%) / (ATR × contract_size)The PPO agent is aggressive (HOLD is often 0%). I cap concurrent positions based on floating drawdown as a safety net against model hallucination.
3 consecutive losses → risk per trade halved to 0.5% and max positions reduced by 50%. Resets after 1 win. This keeps drawdown shallow and prevents over-leveraging during adverse market variance.
If daily loss hits 4%, all trading stops for the rest of the day. Non-negotiable. The model doesn't get a vote.
GOAL: PROP FIRM CHALLENGE
The end goal is passing a proprietary trading firm evaluation. The rules are simple the margin for error is not.
My backtest hit 0.98% max DD with 40.5% win rate well within these limits. Live paper trading is the final validation before attempting the challenge.