Vega Labs

— 01 · Challenge

Researchers wrote strategies in Jupyter, then re-implemented them in C++ for production. Every translation introduced subtle bugs. Backtest results rarely matched live performance.

— 02 · Approach

We built a typed strategy DSL, a research notebook on top of it, and a runner that takes the same DSL to production. One source of truth. Every backtest is a production run with historical data.

— 03 · Outcome

Backtest vs live drift fell from a ~12% spread to under 2%. Time-to-deploy a new strategy dropped from weeks to hours. Researchers stopped duplicating themselves into C++.