The Problem
Manual trading is slow, emotional, and inconsistent. I wanted a system that could scan for momentum opportunities, generate ML-driven signals, and execute trades automatically with proper risk management.
Development Timeline
Built over 2 months (November 2025 – January 2026) with 54 commits. The project evolved through distinct phases:
Phase 1: Foundation (Nov 7)
Initial commit established data infrastructure with SQL Server integration and XGBoost as the ML backbone.
Phase 2: ML Pipeline (Nov 7-8)
Built the training infrastructure: feature engineering, label generation, model trainer. Hit a critical bug early – pandas chained assignment was silently failing to set labels. Fixed by using df.at[actual_idx, label] instead of chained loc/iloc. This was causing 0% positive labels vs expected 63%.
First working model achieved Validation AUC: 0.614. Implemented hybrid adaptive targeting – dynamic profit targets based on ML confidence (50% weight) + volatility/ATR (50% weight), bounded between 1.5% and 5%.
Phase 3: Paper Trading (Nov 9-10)
Launched paper trading with all components wired together:
- AlpacaClient + WebSocket streaming
- PositionManager with persistence
- RiskManager + OrderExecutor
- LiveInferenceEngine for real-time ML
Trailing stop: 1% activation, 0.5% trail. $200k buying power (100k cash + 2x margin).
Bug fixed: Position persistence – system was losing track of positions on restart. Fixed by loading existing Alpaca positions on startup.
Phase 4: Training V2 (Nov 12)
Complete ML pipeline overhaul:
- TickFeatureEngineer with microstructure features (order flow, VWAP, trade size)
- OptimizedLabelGenerator with triple-barrier method
- Expanded from 12 days to 365 days (660M ticks) of training data
- Added news integration for event-driven trading
Phase 5: Real-time Scanner (Nov 17-18)
WebSocket-based momentum scanner detecting rapid price movements:
- 10-second OHLCV bar construction from live trade stream
- Subscribes to ~3,000 symbols with less than 100ms latency
- Slack integration with rich Block Kit formatting
- Claude AI trade reviewer – every trade analyzed by Claude Sonnet 4.5
Phase 6: Scaling (Dec-Jan)
- Wildcard subscription for entire market (~8-10k symbols)
- Database-driven symbol loading
- Windows + Linux deployment via systemd
Architecture
+-------------------------------------------------------------+
| Apache Airflow 2.x |
| +-------------+ +-------------+ +---------------------+ |
| | Training | | Signal | | Position Manager | |
| | DAGs | | Generation | | (Trailing Stops) | |
| +------+------+ +------+------+ +---------+-----------+ |
+----------+----------------+-------------------+-------------+
| | |
+-----v-----------------------------------------+
| MS SQL Server |
| (Historical data, signals, positions) |
+------------------------------------------------+
|
+----------------------v--------------------------+
| Alpaca Markets API |
| (WebSocket streaming + Order execution) |
+------------------------------------------------+
AI-Powered Development
The project extensively used Claude Code for pair programming, with AI co-authoring many commits. The Claude AI Trade Reviewer provides real-time analysis of every trade:
NRXP was bought due to exceptional ML confidence (91.2%), but the extreme volatility (192%) should have been a red flag…
Cost tracking: ~$0.01-0.05 per trade analysis.
Bugs Fixed
| Bug | Symptom | Fix |
|---|---|---|
| Label Generator | 0% positive labels | Use df.at[] not chained assignment |
| Position Persistence | Orphaned positions after restart | Load existing Alpaca positions on startup |
| Indicator Reset | Indicators reset to 0 at date boundaries | Stored procedures with lookback data |
| WebSocket Handlers | Async bar subscription failures | Proper async handler wiring |
| DataFrame Fragmentation | Performance warnings | Use df.assign(**new_cols) |
Results
Current status: Phase 3 live paper trading with 12 symbols. The expectancy journey went from +0.005% (not tradeable) to targeting 0.15-0.30% with news integration. System runs 24/5 during market hours.
Key Learnings
Data quality matters more than model complexity. Spent weeks tuning hyperparameters when the real issue was bad data.
Paper trading reveals edge cases. Backtester did not account for partial fills, after-hours gaps, or WebSocket disconnects.
10-second bars work. Faster signal generation than minute bars, but requires solid infrastructure to handle the data volume.