/research is where you design, test, and refine a trading strategy before attaching it to a live agent. You start by describing what you want the strategy to do — either in plain language or by writing a spec directly — and the builder walks you through backtesting, reviewing results, and iterating with LLM-generated improvement suggestions. When you are satisfied, you create an agent that starts paper-trading your strategy immediately.
Both the Generate path and the Bring your own path produce the same artifact: a StrategySpec YAML that describes your strategy’s indicators, entry and exit rules, regime filter, and parameters. The difference is only in how you author it. Everything downstream — backtesting, the report card, the improvement loop, and agent creation — is identical for both paths.
Builder flow
The first thing you choose is whether your strategy is for Spot or Futures markets. Futures is greyed out as coming soon — spot is the only live mode today.
Making this choice first matters: it means the builder knows your market mode before you write a single line of strategy logic. A strategy that tries to go short, for example, is rejected immediately with a clear explanation rather than surfacing an error later in the process when it is harder to fix.
“Mean reversion on liquid BNB Chain tokens — buy when price dips below its short-term average and sell when it recovers.”
The LLM reads your description and writes a complete StrategySpec YAML. It highlights up to three tunable parameters — the knobs most likely to affect performance — so you know where to focus during the improvement loop. No candle data is needed at this stage; the LLM is working purely from your description and the strategy specification format.
If you already have a StrategySpec YAML, paste it into the editor and skip LLM generation entirely. The builder validates your spec, flags any errors, and proceeds directly to candle import. The improvement loop is optional on this path — you can skip straight to the OOS gate if you prefer.
Both paths produce the same StrategySpec YAML. If you use the Generate path, you can inspect and edit the YAML the LLM produced before continuing. If you use Bring your own, the LLM still reads your spec and can offer improvement suggestions once the in-sample backtest runs — the only thing you skip is LLM authorship of the initial spec.
Select the token pairs and the historical time window you want to backtest against. A progress bar shows the download status for each pair as the data arrives.
These limits keep in-app backtests fast and reduce the noise problem that comes from thin out-of-sample slices. If you request a window longer than 6 months, the builder trims it to the maximum automatically.
The builder runs your strategy against the imported candle data. A per-step progress indicator shows the backtest running. Most backtests complete in under a minute.
The in-sample window is the portion of your historical data used for development. The out-of-sample window — reserved for the prove-it gate in step 6 — is sealed at this point and not used by the LLM or shown to you until the gate.
The LLM reads the in-sample result, identifies the most concrete flaw in the strategy’s behavior, and offers 1 to 3 improvement suggestions. Each suggestion is presented as a structured card:
Each suggestion is framed as a hypothesis — “this should cut chop losses — the backtest will confirm” — not a promise. Pick one suggestion to apply and re-run the in-sample backtest, then review the new report card. Repeat this loop as many times as you want, or skip straight to the out-of-sample gate when you are satisfied.
Every iteration is recorded in the version history with an attempt count visible in the UI — for example, “attempt 7.” This transparency is intentional: you can always see how many iterations went into a strategy’s development, which matters for evaluating how much to trust the result.
The wizard saves your state automatically at every step. A browser refresh will not lose your progress — the builder restores exactly where you left off.
When you are happy with the in-sample result, run the out-of-sample (OOS) backtest. This is a single, one-time run against the reserved portion of your historical data that has been sealed from the LLM throughout the entire development process. The LLM never sees OOS data, which means the improvement suggestions from the previous step could not have been fitted to it.
The report card updates to show OOS metrics side-by-side with the in-sample metrics. This comparison is the honest test of whether your strategy generalizes beyond the data it was developed on.
If the OOS window contains too few trades to draw meaningful conclusions, the builder shows “OOS sample too small to validate” instead of a potentially misleading number. This can happen with low-frequency strategies on shorter time windows — extend your candle import window or try a more active strategy to get a usable OOS result.
The backtest gate is advisory, not a hard block. You can explicitly confirm a failing strategy and proceed to agent creation. The real gate is the manual promote-to-live step: your agent always starts in paper mode and can only move to live capital after you review its forward-test track record and promote it yourself.
Once you are satisfied with the backtest results, click Create agent. Choose an EvaluationConfig preset to define the risk rules your agent will run under:
You can also choose Custom to set every field individually, with inline guidance and a sanity check that warns you if any combination of values is incoherent (for example, a daily drawdown cap set higher than the max drawdown cap).
What the strategy builder does not do
A few things are deliberately out of scope in the current builder:- No mid-loop free-typing. During the improvement loop, the LLM proposes specific changes and you pick one. There is no free-form text box for typing arbitrary edits in the middle of the loop. If you want full control over every parameter, use the Bring your own path and edit your spec directly.
- No self-tuning or automatic optimization. The builder never adjusts parameters on your behalf. Every change is a suggestion you explicitly accept.
- No futures strategies. Spot is the only live market mode. Futures strategies cannot be created or backtested yet.
- No cross-agent comparison. The builder creates one agent at a time. Comparing multiple agents’ forward-test track records is available on the Fleet screen (
/agents), not in the research zone.