BackSpace/LLM backtesting showcase
Before any deep analytics, users get a clear narrative of what changed, where performance held, and why a strategy failed in specific environments. The LLM component translates quantitative backtest output into operator-ready decision context.
> Why did this strategy degrade in elevated volatility during Q4?
The model identifies two drivers: spread expansion reduced fill quality, and entry timing drifted into low-liquidity windows.
Suggested review: tighten liquidity filter, constrain entry to high-participation intervals, rerun by regime segment.
Decision support only. The assistant explains outcomes and tradeoffs; it does not auto-route trades.
I/LLM interpretation layer
Summaries are segmented by trend, compression, and volatility expansion regimes.
The assistant explains why sizing, slippage, or entry rules altered outcomes.
Narratives tie back to run metadata so every statement is traceable to a result set.
II/Backtest compute foundation
Equity & drawdown
Monthly returns
By environment
Sharpe
1.84
Profit factor
2.31
Max drawdown
-11.4%
Win rate
58.2%
Run the decision against the exact market context it originally faced.
Break performance out by regime to identify where strategy edges collapse.
Each run keeps its seed, params, and source revision for exact reruns.
III/Dashboard direction
Dashboard component plan · LLM backtesting assistant
The tab is positioned as an operator assistant: ask why drawdown clustered, compare two runs, and get suggested next experiments before re-running compute.
It is built as a modular UI surface so you can fork the model provider later without redesigning the surrounding dashboard workflow.
Current product stance: explain and prioritize decisions, never auto-execute positions.
Prompt and response surface designed to be provider-agnostic.
Assistant phrasing assumes run ids, environment slices, and model metadata.
No direct trade routing from assistant output.