Running backtests
Run a saved strategy revision against selected historical bars and inspect its artifacts.
A Lab backtest associates a source revision with historical data and saved results. Use a short baseline run before asking the assistant to perform repeated experiments.
Check the settings
In Settings, review the instrument, UTC start and end, interval, timezone, capital, quantity, fill delay, regular-hours option, and end-of-day closing. These choices affect the interpretation of the run.
The current Lab supports bar-based experiments. Tick, order-book, book-heartbeat, and interpolation configurations are unsupported. Polymarket imports are also unsupported by this workflow.
Start and follow a run
Choose Run backtest. The Lab saves pending changes first and attaches the run to that revision. Select the run in History to see its status and results. Use Cancel run in Chart to stop an active run.
The experiment permits one active backtest and one active agent task. The UI prevents starting a separate manual run while an agent task is active; the agent can run tests as part of its own task.
Leaving the page does not cancel work while the backend remains running. Returning restores saved experiment state. Backend restarts do not resume queued or running jobs automatically; check their status before retrying.
Understand reused data
After the first successful run for a data selection, the experiment reuses its saved bars for that selection. Changing source can therefore compare behavior against the same observations.
The selection includes instrument, interval, start, end, and timezone. Execution settings such as capital and fill delay still need separate comparison even when the dataset matches.
Read completed artifacts
| View | What to inspect |
|---|---|
| Chart | Historical bars, plotted output, and annotated regions |
| Trades | Returned trade records and their context |
| Metrics | Selected run and pinned baseline summaries |
| Logs | Saved strategy diagnostics |
The trade table displays the first 1,000 records; the assistant can page through the full result. Completed-run logs are stored separately and loaded in pages. Use the time and text filters to focus an investigation before their retention period expires.
Runs are limited to 100,000 bars and bounded result size. Narrow the period or reduce output when a limit is exceeded. For price checks, delays, warmup, and result interpretation, read the backtesting guides.