- ~30% average returns
- 24/7 live in production since 2024
- tick-level feature engineering
A transformer that predicts short-horizon crypto movements, trained on Google Cloud, with features engineered from high-frequency tick data. Raw price is mostly noise, so a lot of the work went into signal-processing methods to denoise it before the model ever sees it.
Predicting well and making money are different problems, and the gap between them is infrastructure.
The pipeline
- Backtesting that replays real market conditions: fills, fees, slippage. Assuming you transact at the mid price is the fastest way to backtest a strategy into profits that vanish live.
- Walk-forward validation, so a strategy has to survive regimes it never trained on before it touches real capital.
- Live execution through the Binance API from an AWS EC2 instance, running 24/7.
- Reliability work: reconnection handling, position reconciliation, fail-safe behaviour. A crash mid-position is not a stack trace, it is money.
Averaged around 30% returns, optimising for risk-adjusted performance rather than peak headline number. Running and being iterated on since July 2024.
Private repository. Happy to walk through the architecture and backtesting methodology.