Skip to content

AI Chat

AI Chat is one assistant for everything you do before risking capital: ask about markets, research a question on the web, build or refine a strategy, and backtest it, all in the same conversation. There's no separate "trader" or "builder" mode to pick: just type.

The same AI Chat is in the Live section too (Trader and Quant), in Live's blue colours. It's the same assistant with the same conversations: switching between Testing and Live keeps the chat you have open.

Not financial advice

AI analysis can be incomplete or wrong. Every answer is a starting point for your own judgement, not a recommendation to trade.

Ask about markets and your work

Open AI Chat from the sidebar and ask a question, for example:

  • What's moving in crypto today?
  • Review my strategies.
  • Why did my last backtest lose money in March?

On the Trader and Quant plans the assistant can read your own work (saved strategies, backtest results, paper runs) and current market context, and it shows where each fact came from under the answer. On Free, questions are answered with web research only (below).

When a reply ends with suggested follow-ups, tap one to send it as your next message.

Monte Carlo and prop-firm checks

On Trader and Quant, ask about the risk of a finished backtest, for example "Run a Monte Carlo check on my last BTC backtest" or "Which prop firms would this backtest pass?"

  • Monte Carlo reshuffles the backtest's trades into 2,000 paths. The answer quotes:
  • the risk score;
  • the chance of a drawdown deeper than the ruin threshold;
  • the median worst drawdown;
  • the worst drawdown in dollars that only 1 path in 20 exceeds.

The dollar figure uses the backtest's starting capital unless you name an account size. Reshuffling keeps the final return fixed, so the chance of profit isn't estimated here. The Risk page's other methods estimate it. - Prop firms replays the backtest against the prop-firm catalog. The answer lists the best 10 programs, each with pass, open or fail and the first rule that failed. It also says how many programs passed, are still open and failed. - Full results, charts and settings are on the Risk page, linked under the answer. Like the Risk page, checks are limited per minute. They use no AI credits beyond the answer itself. - These checks only read results. They never start a run or make a plan.

Research on the web

Ask for news or facts and the assistant can search the public web and read a few pages before it answers.

  • Cited sources. Facts from the web carry markers like [W1]. Select one to jump to its source card, which shows the page title, the site, when it was published (or "Publication date unknown") and when it was retrieved. Cited sources come first. Pages it looked at but didn't cite are under Also retrieved. Some primary sites (governments, central banks, regulated exchanges, company investor relations) are labelled Official source.
  • Links you paste. Paste an https:// link and ask about it; the assistant reads that page directly.
  • Per answer: up to 2 searches and 3 page reads. An answer that uses web evidence costs one extra AI credit. A search that finds nothing costs no credit but still counts toward your monthly searches.

  • Monthly search allowance (resets on the 1st, UTC):

    Plan Web searches per month
    Free 30
    Trader 150
    Quant 500

    When 10 or fewer are left, the remaining count appears under the sources. When none are left, the assistant says so and when the allowance resets. Reading links you paste still works. Occasionally search is busy for everyone; try again later.

What it will never do:

  • Invent a source. If search is unavailable or finds nothing, it says so. A marker that doesn't match a retrieved page is flagged as unverified.
  • Open private or internal addresses. It only reads public https:// pages from its own search results or from links you pasted.
  • Send your private data to the search engine. It refuses searches that repeat your strategy names or account figures from the conversation.
  • Follow instructions written on a web page. Page text is treated as information. Nothing on a page can start, change or accept anything; plans still need your acceptance.

Build and refine a strategy

Describe an idea in plain English (an entry rule, a market, a timeframe) and the assistant asks about anything missing, then writes the strategy when you select Generate code. Keep chatting to change it; a research question in between doesn't disturb your draft. See Getting started for the full walkthrough and Strategy authoring for writing strategies by hand.

Improve a saved strategy

Choose Improve in chat on a strategy in your Library or on a backtest result, or pick it with Strategy in the message box. Then say what you want, for example "Lower the drawdown without changing the entries." The assistant agrees a short brief with you:

  • Goal: what the change should achieve.
  • May change: which parts the assistant may touch. Choose any of Parameters, Entry & exit rules, Filters & indicators and Risk & sizing.

Select Generate candidate. The assistant writes a new version, a candidate, and your original strategy stays exactly as it was. A candidate is created only when its code passes validation and stays inside the brief. For example, if Risk & sizing isn't ticked, the candidate's RISK block can't differ from the original's.

The candidate card shows:

  • the assistant's list of changes and why;
  • the code changes, compared with the version it came from or with your original;
  • Run backtest, Add to Library and Apply to original.

Keep chatting to refine it further: each new candidate builds on the last one, and all of them stay listed under the original on its Library page. Regenerate on a candidate's reply makes another candidate from the same starting point; the earlier one stays in your Library. Candidates join the original's family, so the Optimizer Lab counts their trials together.

Apply to original replaces your original's code with the candidate as it is now, including any edits you made to it in the Editor. It works only if the original's code hasn't changed since the candidate was made, or if it holds exactly the earlier candidate this one builds on (because you applied that one). Renaming the original or adding it to the Library doesn't count as a change. A strategy built without code can't be replaced: add the candidate to your Library instead. If the original is read-only, changing its code needs Trader or Quant.

The server can tell parameters, risk settings, timeframes and the rest of the code apart, but it can't tell entry and exit rules from filters inside the same function. When you allow one but not the other, the assistant follows your brief, but the server can't check it.

When generated code fails validation

Every generated strategy is checked before you can use it. If it still fails after the assistant's repair attempts, the reply shows Failed validation with the problems found. You can't save or run it; select Fix it to have the assistant repair it. A reply that writes no working strategy, or the same code again, is labelled too. None of these use AI credits (up to 10 a day). If the checker itself was unavailable, the code is labelled Couldn't be checked and you can still use it.

Bring your own code

  • AlphaProve Python: paste a strategy written for AlphaProve (it has HISTORY = … and def evaluate(ctx)), or select Open file to load a .py file. It's checked and opens as a new draft, without using AI credits. Importing your own Python needs Trader or Quant. To ask about code instead ("Why does this never trade?"), put the question above or below it: the assistant answers and nothing is imported.
  • TradingView Pine Script: see Pine Script.
  • Other platforms: MetaTrader (MQL4/MQL5), cTrader, NinjaTrader, backtrader, freqtrade, zipline, QuantConnect and EasyLanguage code can't be imported yet. Describe the strategy's rules instead and the assistant builds it.

Backtest and forward-walk plans

Ask the assistant to backtest one of your saved strategies, for example "Backtest my RSI strategy on BTCUSDT 1h for the last 6 months with 10,000 USDT at 1×." It answers with a Backtest plan card. Nothing runs until you accept it.

The card shows exactly what will run:

  • the strategy's name and a short revision label (the full revision is under Exact revision);
  • the market, timeframe and period, for example BTCUSDT · 1h · Mar 30, 2026 → Sep 29, 2026 · 10,000 USDT · 1×;
  • that leverage comes from the plan and replaces the strategy's own, while position sizing and other risk rules come from the strategy's RISK setting;
  • when the plan expires ("expires in 23 h"). Plans are valid for 24 hours.

Plans must fit your plan's limits (history length and maximum leverage) and the available market data; otherwise the assistant tells you what to change. A plan can't be edited. If you want different settings, ask for a new plan.

  • Accept and start starts the backtest at exactly those settings. It uses one backtest from your daily allowance, like any other run.
  • Cancel plan throws the plan away. It never stops a paper-trading run.

If the strategy's code or parameters change after the plan was made, or the plan expires, the card says so and Accept is disabled. Ask for a new plan. Renaming the strategy or adding it to your Library doesn't affect the plan.

If the strategy isn't in your Library, the card says it will be added; Accept adds it, so its runs show there.

Forward walks

Ask for a forward walk the same way, for example "Run a forward walk on my EMA strategy, BTCUSDT 1h, March to July, 2 months in-sample and 1 month out-of-sample." The assistant answers with a Forward-walk plan card. To test the strategy on the full window first, ask for a backtest and then a forward walk. You get one Backtest + forward-walk plan card with matching symbol, timeframe, window, capital and leverage, and one Accept and start runs both, in order. If the backtest fails, the forward walk doesn't start.

  • The strategy must declare parameters to sweep: a sweep list in PARAMS (see Parameters & forward-walk) or on a numeric no-code field. The walk uses that parameter grid, as on the Forward Walk page: up to 150 combinations on Trader and 200 on Quant.
  • Forward walks need Trader or Quant, and you can run one at a time. If another forward walk is queued or running, the step waits for a free slot (until the plan expires).
  • A walk can run for up to 1 hour on Trader and 3 hours on Quant. While it runs, the task card shows roughly how long is left; live batch progress is on the walk's own page.
  • Forward walks don't use your daily backtest allowance. In a combined plan, the backtest does.

Optimizer Lab campaigns

Ask for a Lab campaign by name, for example "Run a quick Optimizer Lab campaign on my EMA strategy, BTCUSDT 1h, from January 2025 to yesterday." Words like optimize or tune the parameters work too. Asking to improve a strategy makes a candidate instead (see above). The assistant answers with a Lab campaign plan card:

  • the parameters the Lab will tune and the ones it keeps fixed, and the estimated number of combinations;
  • the depth, its estimated duration, and its cost in AI credits: Quick 10, Standard 20, Deep 60;
  • how many Lab campaigns you have left this month;
  • a warning when earlier campaigns on this strategy already looked at the same final test period. A result that reuses it is not a test on unseen data.

Accept and start submits the campaign exactly as the Optimizer Lab page would, with the same plan limits, monthly quota and checks.

  • Lab campaigns need Trader (Quick and Standard, 8 a month) or Quant (all depths, 20 a month, up to 8 of them Deep).
  • The period must cover at least 600 days. The strategy needs 1 to 8 tunable parameters in PARAMS that give at least 50 combinations.
  • From chat, a no-code strategy stays no-code and the AI signal probe is off. Use the Optimizer Lab page for either.
  • You can run one Lab campaign at a time. If another one is running, the step waits for it, until the plan expires.
  • The credits are reserved when the campaign starts. If nothing clears validation, you get half back. A failed campaign is refunded in full. A cancelled one is charged only for the part that ran.

When the campaign ends, the task card shows the Lab's verdict and links the report. Ask AI to analyze explains it.

Compare two versions

To test a candidate against its original, or any two of your saved strategies, ask for a comparison and say what "better" means, for example "Compare the candidate with the original on BTCUSDT 1h for the last 6 months: higher Sharpe, and max drawdown must not get worse." The assistant answers with a Comparison plan card. It shows both strategies, the shared settings and the goal in words, such as Goal: higher Sharpe; max drawdown must not get worse.

  • Both strategies are backtested on identical market, timeframe, period, capital and leverage. The comparison uses two backtests from your daily allowance.
  • If you also ask for forward walks, both strategies get one. Their results are shown beside the verdict for information; they don't change it.
  • Position sizing still comes from each strategy's own RISK setting. If the two differ, the card and the result say so.
  • Accept adds both strategies to your Library, like any plan.

AlphaProve computes the verdict from the stored results. The AI doesn't choose it.

Verdict When
Candidate better The goal metric improved by more than the minimum difference, consistently across the period, and the metric you said must not get worse didn't.
Candidate worse The goal metric got worse in the same way, or it improved but the metric you protected got worse ("improved Sharpe but worsened max drawdown").
Inconclusive Fewer than 30 trades on either side, a difference below the minimum, or a difference that doesn't hold across the period.
Not comparable A backtest didn't finish, or the two didn't run on the same settings, for example because a strategy was deleted or edited after you accepted.

The minimum differences are: - Sharpe 0.10; - return 1 percentage point; - max drawdown 1 percentage point; - profit factor 0.05.

For Sharpe, return and profit factor, the period is split into 12 equal slices and the improvement must hold across them. A max-drawdown goal uses the minimum difference only. Every result also says two things: - how many versions of this strategy you have tested: the more you try, the likelier one wins by luck; - that the period is not new data for a version designed after looking at it.

Task cards

Once accepted, the plan becomes a task card in the conversation. The task runs on our servers: you can keep chatting, close the tab, or come back later on another device and the card shows where it got to.

Status Meaning
Starting The task is about to start a run (after Accept or Retry).
Running A run is being started (or its start is being retried after an interruption).
Waiting for results A run is in progress. Results appear on the card when it finishes.
Waiting for capacity All your run slots are busy; it starts when one frees up (until the plan expires).
Needs your input Something changed since you accepted, such as the strategy's code or parameters were edited, your plan's limits changed, or the plan expired. Ask for a new plan or cancel.
Completed The headline results are on the card (for a backtest: return, Sharpe, max drawdown).
Failed The card says why in plain words.
Cancelled You cancelled the task.

A forward-walk step shows Waiting for a free forward-walk slot while another walk is queued or running, then Running with about how many minutes are left. For a finished walk, the card shows its out-of-sample return, walk-forward efficiency, max drawdown and number of batches.

A Lab step shows Waiting for your other Lab campaign to finish while another campaign runs, then Running with a link to the live campaign. When the campaign ends, the card shows: - the verdict; - the number of trials and the checks passed; - AI credits used against the credits you accepted.

View report opens the Lab report. Cancel task cancels the campaign; its report so far is kept.

When a comparison finishes, its card shows both results side by side: each metric, the differences, the verdict and its reason, and links to both backtests.

  • View result opens the full result: a backtest's tearsheet in Results, a forward walk on the Forward Walk page. In a backtest + forward-walk task, each step has its own Result link.
  • Ask AI to analyze asks the assistant what the result means.
  • Cancel task stops further runs and asks its running backtest to stop. A backtest that is still loading market data can take a few minutes to stop. A running forward walk stops at its next batch, which can also take a few minutes. Results so far are kept, and paper-trading runs are never affected. If you cancel before the backtest starts running, it's given back to your daily allowance: straight away with Cancel task, or within a few minutes if you cancel it from Results. A backtest that already started, finished or failed still counts. Backtests you start outside a chat task (from Results, the Library or Run) keep their charge when cancelled while queued.
  • Retry appears only when a run was interrupted on our side (for example, a server restart). It starts one new run with the same settings and uses one more backtest from today's allowance. Nothing was charged twice for the interrupted run. For a forward walk, Retry starts one new walk with the same settings.
  • If our server stops in the middle of a forward walk, its step keeps showing Running until our clean-up marks it as interrupted. That can take up to about 3 hours 20 minutes after the walk started (the longest a walk may run, plus a margin). Then Retry appears.

Notifications

When a task finishes, fails or needs your input, a notification appears on whatever page you're on, naming the conversation and the result. On a computer it sits at the top right on chat and Terminal pages, and at the bottom right elsewhere. On a phone it appears at the top of the screen. Up to three show at once; the next one appears when you clear one.

  • Open chat takes you to the conversation. Dismiss clears it here, and on your other open tabs at their next check.
  • Notifications stay until you act on them. If the tab was in the background, they appear when you come back to it.
  • You don't get one for the conversation you're looking at, or for a task you cancelled yourself.
  • A task notifies once per outcome: if it fails again after Retry, there's no second notification.

Stop and retry a reply

While the assistant is writing, Stop response ends that reply only. It doesn't cancel a task card. A stopped reply shows Response stopped. If it stopped before writing anything, Retry writes it again in place. For a partly written reply, use Regenerate under the reply.

History

Chat History lists every conversation in one place, including conversations from the former AI Trader. Old AI Trader links open the same conversation in AI Chat.

  • Filter by what happened in a thread: Research & analysis, Gather, Generate, Backtest review, Pine import, Pine export or Idle. Filtering searches older conversations too.
  • Pin, rename or delete a conversation. On a phone, these are under More actions.
  • Export as Markdown downloads a readable copy of the conversation: the messages, a short list of the tools used, the strategy and backtest analysis, and the sources with their dates. Images and links written by the assistant are exported as plain text.
  • Export with tool data downloads the same copy plus the raw tool inputs, outputs and structured results as JSON, in a separate file ending in "tool data".

Limits

  • Answers use AI credits from your plan's monthly quota (your remaining credits are shown under Settings). Backtests, including accepted plans, use your daily backtest allowance. Forward walks don't, but you can run only one at a time.
  • A Lab campaign from chat uses AI credits like one started on the Optimizer Lab page, and counts toward the same monthly campaigns.
  • When the assistant only asks you to say what you mean, that reply is free (up to 20 a day). Generations that end with failed or no usable code are free too (up to 10 a day).
  • The assistant only suggests paper trading. It never starts, stops or changes a paper run by itself, and a plan never touches one.
  • AI analysis can be incomplete or wrong. Not financial advice.