Back to the projects
Own projects · Work in progress · simulation only
DealBot – a trading system for learning
My own project since 2026: a system that tests trading strategies and machine learning models on real market data and runs them as paper trading bots – with simulated money only. It is about learning: data pipelines, clean backtests, comparing models fairly and building an application end to end.
Design mockup with invented example numbers – not real results.
Building blocks
- Data pipeline: prices, news, index lists (e.g. S&P 500 members at any point in time) and exchange rates from several sources, with cross-checks
- Backtester with realistic trading costs
- Paper trading bots: signal + risk profile + protective stops + capital
- A scanner that suggests new candidates – the decision stays with a human
- Web dashboard with bots, charts, decision log and inbox
The model tournament
- Rule-based strategies (e.g. trend following, mean reversion, dual momentum) and ML models (e.g. LightGBM with and without news, LSTM) compete on the same data
- Walk-forward: always train on the past and test on the period after – after costs
- Ranked by risk-adjusted performance, not by profit alone; a model that allows too large losses is out
- “With / without news” pairs show whether the news mood really helps
Safety first
- Everything runs as a simulation (paper trading); real trading is locked
- Kill switch for all bots
- Access keys are write-only and never shown again
Tech
- Python, FastAPI, pandas, LightGBM, PyTorch
- React and TypeScript for the dashboard
- Docker on my own server; heavy training on my own graphics card