Two Backtest Engines, One Pipeline
Search and validation want opposite things: search wants throughput across a huge parameter space, validation wants full realism on a handful of survivors. We run two purpose-built engines instead of compromising on one, and combine them for anything complex enough to need both.
Vector Engine
One parameter grid, run at once — not bar by bar.
The whole strategy is expressed as array operations and compiled kernels — there is no per-bar loop at all. Best for trend-following scenarios and any coarse parameter sweep: thousands of indicator/threshold combinations across the full history, ranked in seconds. Its job is not the final answer — it is to cut a huge search space down to a short-list of candidate configurations and the market slices (regimes, time windows) worth a detailed look.
Faster than open-source engines
One identical parameter sweep — 150k bars, 80 indicator combinations, trade count locked to the same value — how many times faster our GPU engine runs it than each open-source engine.
Event Engine
Bar by bar, tick by tick — every fill earns its place.
A per-bar / per-tick replay against trades and full order books, with candles, footprint (order-flow per price level) and volume profile layered in as needed. This is where a strategy meets realistic fills, slippage, margin and queue position — and where an AI or news-analysis agent gets to make one decision per candle, exactly as it would live. It is also the only engine we trust for HFT strategies, where the order book and latency are the strategy. Slower by design: that per-event cost is the price of realism, and for a short-list of survivors it is negligible.
Further reading
Questions about the engines?
Talk to the team — we usually reply fast.