Back to selected work
Real-time signal engine
A BTC prediction-market engine and terminal dashboard, from feature pipeline to execution simulator.
Outcome
A validation pipeline that reports 48 to 50% out-of-sample instead of claiming an edge.
- Year
- 2026
- Role
- Design and build
- Status
- Private, in progress
- Stack
- Python, LightGBM, FastAPI, React, WebSockets
The problem
Noisy one-second market data, thin edges, and execution assumptions that flatter a backtest.
What I built
A Binance spot stream, one-second features, a LightGBM probability of up, a decision state machine, a maker-only execution simulator with fees and liquidity gates, outcome buckets, and a WebSocket view.
Pipeline
- Binance stream
- 1s features
- LightGBM
- Decision state
- Maker simulator
- WebSocket view
Validation
CPCV
PBO
DSR
Leakage audit
Triple-barrier labels
Calibration
Result
48 to 50% out-of-sample on spot models. I report it as a measurement, not an edge. The useful work is the validation pipeline and the per-bucket record.