BTC paper-trading system

A model with a good backtest is easy. A system that ingests live data, retrains as the market drifts, sizes positions, pays realistic fees and logs every decision — across two machines — is the actual engineering problem.

• Market data ingestion, feature engineering, ML signal generation and simulated order execution with full trade logging, as one pipeline.
• Market-structure features — order blocks, EMA pullback breakouts, liquidity sweeps — feeding a LightGBM classifier trained on 90 days of Binance OHLCV data.
• An adaptive engine with rolling retraining to counter model drift.
• A backtester that models exchange fees, position sizing and risk/reward geometry.

• Fee modelling and position sizing live inside the backtester, not as an afterthought, so simulated returns are comparable to what execution would produce.
• Compute (macOS host) and execution (Windows gateway) are separate processes on a private Tailscale mesh, which keeps the execution side simple and isolated.
• Traced and fixed stale-data, duplicate-process and fee-drag defects that had been inflating simulated returns — the fixes lowered the numbers and raised the trust.

• A LightGBM signal with 0.567 AUC on the held-out window, retrained on a rolling basis.
• A fully logged paper-trading loop running unattended across two machines.
• Three classes of silent return-inflation bugs found and closed.

• Promote the strategy evaluation into the pre-registered crypto-ml-study workflow.
• Add regime detection to gate signals during low-liquidity periods.
• Replace ad-hoc alerts with a small dashboard of live positions and drift metrics.