Building by chat
The chat dock is how you edit a strategy without writing JSON. You describe a change in plain English; the assistant rewrites the strategy spec, explains what the change means in trading terms, and re-runs the backtest when it makes sense to.
What the assistant can do
- Edit the spec. Every edit comes back as a complete updated strategy, not a vague suggestion — you see exactly what changed, and the run happens on the real, validated spec.
- Explain historical results. It has your latest run in front of it — returns, drawdown, trade counts, exit breakdowns — and can tell you what drove them.
- Teach concepts. Ask what Sharpe means, what an ATR stop is, why drawdown matters. Jargon gets explained, not assumed.
- Compare runs. Ask how the current version stacks up against the previous one and why.
When a request is vague — “make it better” — it proposes one concrete change with reasoning, not a rewrite. The more specific your ask, the better the edit: “tighten the stop from 8% to 5%” beats “tighten the stop”.
Prompts that work well
| You want to… | Try saying |
|---|---|
| Change a parameter | “Tighten the stop from 8% to 5%.” |
| Add an entry filter | “Only enter when the 14-day RSI is above 55.” |
| Rename the strategy | “Rename this to Monday Reversal.” |
| Change the timeframe | “Switch this to 15-minute bars.” |
| Understand a result | “Why is the max drawdown so high?” |
| Test a hunch | “What if I stayed long all year instead?” |
| Learn a concept | “Explain what Sharpe ratio means here.” |
Two conveniences worth knowing. First, simple indicators referenced in a rule are registered automatically — mention rsi_14 or sma_50 in an edit and the matching indicator definition is added for you. Second, when you switch to an intraday timeframe the assistant also shrinks your date range to 60 days or less and tells you why — that cap is real and explained in Markets & timeframes.
Starting from scratch
With no strategy loaded, the chat switches to a guided intake. It asks one question per turn, in a deliberate order:
Market
US stocks and ETFs, or crypto via Polygon-style tickers likeX:BTCUSD.Timeframe
Daily bars, or intraday (60m down to 1m — with the 60-day history cap flagged up front).Entry idea
What gets you into a position — a crossover, a breakout, a seasonal window, whatever you have.Exits and stops
What gets you out, and where the stop-loss sits.Sizing
Percent of equity per position (or percent risked per trade), and how many positions at once.
After every answer you see a full draft of the spec, live. Parts you have not answered yet are filled with clearly-labeled provisional defaults — SPY, daily bars, 10% sizing — so the draft is always complete and runnable. Nothing runs until you say the strategy is ready to test.
Warnings it raises on purpose
Whenever it edits a spec or summarizes results, the assistant scans for a short list of well-known ways backtests mislead, and warns you — then carries on. It never blocks your edit; it makes sure you know the risk you are taking with the analysis, not the trade:
- Look-ahead bias— rules that peek at data the trade could not have known yet, like entering at the open using the same bar’s close. The backtest looks brilliant; a real trader could never have done it.
- Overtight stops— a stop far smaller than the timeframe’s typical bar-to-bar swing mostly buys noise exits, not protection.
- Unrealistic sizing — near-100% of equity in one position, or position size × max positions far beyond 100% (implicit leverage).
- Survivorship bias— hand-picking the famous winners of today bakes tomorrow’s knowledge into yesterday’s universe, flattering any strategy tested on them.
- Too few trades — below roughly 30 trades, the stats are anecdotes, not evidence.
What it refuses — and why that is a feature
The assistant will not, under any phrasing:
- predict future prices,
- recommend buying or selling any real security with real money,
- claim a strategy “will” make money.
Ask anyway and you get a polite refusal plus a useful counter-offer: “I can’t predict prices, but we can test how that idea would have performed historically.” There is also nothing behind the chat that could act on advice if it gave any — the product has no broker connection and places no trades, ever.
Guardrails, and the one retry
Every spec the assistant proposes is validated on the server before it touches your strategy: numeric fields are clamped to documented bounds, tickers are checked, and every rule expression is parsed against the engine’s grammar. If validation fails, the errors are sent back to the model for exactly one self-correction attempt. If the retry fails too, the edit is dropped — you see the reply and the validation errors, and your current spec stays untouched. A failed edit never half-applies.
What chat costs
Nothing per message — AI chat is included on every plan, with no counter and no per-message charge. What exists instead is a daily limit: 3 messages without an account, 5 on Free, 15 on Pro, 40 on Max. Model tokens are one of the two things that genuinely cost us money per use, so the free tier’s limit is a real constraint rather than a formality — worth knowing before a long session. Context is capped on every plan too (the last 12 messages, each trimmed to 2,000 characters), so a long conversation can’t quietly balloon. Full table in Plans & limits.