Introducing the AnalyticsFX AI research assistant and the Builder

Most trading ideas start as a sentence. "The euro follows the two-year rate spread a few days later." "A London breakout with ATR stops still works on GBPUSD." Turning that sentence into something you can check usually takes a data download, a spreadsheet, an evening of code and a lot of patience. Today we are releasing two tools that shorten that path: the AnalyticsFX AI research assistant and the Builder.

The research chat

Open AnalyticsFX AI from the dashboard sidebar (/dashboard/research) and ask the way you would ask a colleague. There is no query language to learn and no ticker format to remember.

Answers with sources

The assistant searches a research library of more than 40,000 documents. It holds working papers from the BIS, the ECB, the Federal Reserve, the Bank of England, the Riksbank, the SNB, the IMF and arXiv quantitative finance, plus central-bank communication: FOMC statements and minutes, ECB press conferences and about 21,000 speeches from roughly 90 central banks.

Each claim in an answer carries a numbered marker, and the Sources panel lists the documents behind those markers with their title, authors, year and institution. A marker that does not match a document the assistant actually retrieved is removed, so a number in the text always points at something it read.

The library has gaps, and the assistant says so: journals behind a paywall are not in it, and nothing from the last few weeks is. For today's news it can search the web and shows the pages it used. Web searches come with a monthly allowance on the paid plans: 50 on Core, 200 on Terminal and 1,000 on Pro.

It reads the market, and your own trading

While it answers, the assistant pulls the data it needs from AnalyticsFX: prices on every timeframe, COT positioning, seasonality, the economic calendar with its surprises, central-bank rates and news. When a question is about you, it reads your journal, your backtests and your forward trades, and only ever from your own account. Each data set keeps the plan it has everywhere else on AnalyticsFX.

Dashboards built from real candles

Ask for an overview and the assistant can lay out a dashboard of live cards: numbers, bars, lines, pies, scatter plots and candles. Every figure is computed from the bars in our database rather than typed in by the model, and a figure the model only stated is marked "Stated, not measured". Refresh the figures to read them again from the latest bars, or save the dashboard to a project as a single HTML page.

Hypotheses, tested out of sample

When a paper makes a claim you could trade, the assistant can write it down as a hypothesis card before any test runs, so the claim cannot be bent to fit the answer afterwards. In the research lab you run the test: models are fitted against two simple baselines with walk-forward validation, and the result is shown out of sample next to in sample, with the result of every fold and a check on other markets.

The verdict is one of four: Holds up, Refuted, Overfitted or Inconclusive. Every look at the data is counted, and the bar for significance rises with the number of tests you run, so a lucky run does not pass for an edge.

Backtests in your browser

Strategies from the chat run in your browser, with the trades drawn on the chart. You can run a batch of up to 30 variants and rank them side by side. Code you have not checked yet runs in an isolated sandbox without network access.

The Builder

The Builder is where your indicators and strategies live: a code editor, a live preview and an AI assistant side by side, with your indicators and your strategies on two tabs.

Describe it, write it or import it

• Describe it. Ask the assistant beside the editor for a change, or ask the research chat for an indicator or a strategy. It writes a first version, checks that it compiles and opens it in the Builder as a draft.
• Write it. Indicators and strategies are written in TypeScript. Sessions and time zones, several timeframes and markets, calendar surprises, rate differentials and COT are available as inputs, along with stops and targets.
• Import Pine. Paste a TradingView indicator or strategy and it is converted into TypeScript. The converter targets Pine Script v5. Whatever it cannot convert stays in the code as a TODO comment, so nothing disappears silently.
• Switch kinds. Ask the assistant to turn an indicator into a strategy, or a strategy into an indicator, and it creates a new draft of the other kind.

Drafts are free to build and try. A draft is deleted 14 days after its last change unless you save it, and every save keeps a version you can restore. Only code you have checked runs in the preview, so press Check code first.

Backtest with your own settings

Pick the market, the timeframe and the period (Recent, 5Y, 10Y, 20Y or Max), and override the starting balance, the spread, the commission per lot and the leverage for that run. These settings apply to the run only and never change your file. The result shows net profit, return, max drawdown, profit factor, win rate, number of trades, costs and Sharpe, with the equity curve and the list of trades.

Optimise without fooling yourself

The Optimise tab has three modes:

• Optimise. Give each input a range (from, to and step) and search it as a grid, at random, or at random and then refined around the best result. Rank by net profit, Sharpe, profit factor, drawdown, win rate, return over drawdown or an objective of your own. Only the first 70% of the period chooses the settings, and the rest is held back to check them.
• Walk-forward. Optimise on one stretch, test on the next, and repeat across the history in rolling or anchored windows. The out-of-sample stretches are joined into one equity curve, and the verdict says whether the strategy held up out of sample.
• Other markets. Run the same rules on up to 12 more markets, to see whether the edge travels or was fitted to one chart.

Debug, Profile and modules

The debugger steps through a run bar by bar, and the profiler shows where the time goes. Modules are code you save once and import into any script, so a helper you trust lives in one place.

Alerts, publishing and going live

• Alerts. Call alert() in a script, and the Alerts tab runs it on live bars and tells you when it fires.
• Publish. Publish a copy to the shared library, where anyone signed in can read it, clone it and comment on it. A strategy carries the grade and stats of its last valid backtest.
• On your chart. Add a saved indicator to the live chart in one click.
• Go live. A saved and checked strategy can go live and run on our servers. It emails you each signal, or trades a forward testing account with its own stop and target.

What each plan includes

The research chat and drafting in the Builder are open on every plan, Free included. From there:

• Core saves up to 5 indicators and 3 strategies, and adds strategy backtests, the debugger, the profiler, modules, research projects and shareable conversation links.
• Terminal adds the research lab, Optimise with walk-forward and other markets, and comparisons across several markets. It saves up to 20 indicators and 10 strategies.
• Pro adds publishing to the library and up to 3 live strategies, and saves up to 50 indicators and 25 strategies.

Plans are priced in euros, including VAT: Core costs €14.99, Terminal €34.99 and Pro €49.99 a month, with yearly billing as an option. The plans page compares every limit side by side.

How AI credits work

Every answer uses AI credits. What a message costs depends on how much research it takes, and you see the cost before you send. The assistant in the Builder draws from the same balance.

• Free: 60 credits once, after you confirm your email.
• Core: 500 credits a month.
• Terminal: 2,000 credits a month.
• Pro: 5,000 credits a month.

The monthly allowance renews every month and does not roll over. If you need more, credit packs top up your balance at any time: 1,000 credits for €9, 5,000 for €39 or 20,000 for €129. Bought credits never expire. Running a hypothesis test on our servers costs 25 credits, while re-running it in your browser is free.

What it is not

The assistant is a language model. Its answers can be incomplete or wrong even when they cite a source, so open the source before you rely on it. A backtest, an optimisation or a hypothesis that holds up describes the past, and none of them promises future results. Treat everything here as research, not investment advice. Every trading decision stays yours.

How to start

Open AnalyticsFX AI from the sidebar and ask your first question, or open the Builder and start a new indicator. For an overview of everything the assistant can do, see the AI research page.