From an idea to a tested strategy, step by step.
Here is everything Takumis does, in the order you would use it. All of it works in the private beta today, where Takumis takes a strategy as far as a finished, tested backtest. What we are building next is listed at the end.
Describe
Say what you want, in your own words and your own language.
Strategy building in plain language
Describe the rule (entry, exit, size, filters) and get the whole strategy back as rules you can read, in plain language and as a flowchart. Every number you say becomes a value you can change by asking.
Market analysis in plain language
Ask about a market move and Takumis looks at the price action and the economic calendar together, so "what happened to EURUSD on Thursday" comes back with the releases and price moves around that day. It shows you what lines up, though no one can say for certain what moved a market.
Explanations at your level
Choose beginner, intermediate or expert explanations, and change it any time in settings.
Speak instead of typing
Talk to Takumis instead of typing. Voice input works in more than 100 languages and recognises which one you are speaking.
Over 50 indicators
From moving averages to volatility, momentum and volume indicators.
Works on your phone
The whole app works on mobile, not only on a desktop.
Test
Ten years of real market data, with spread, commission and swap applied.
Backtest on ten years of data
Minute data from May 2016 to today, on 56 currency pairs plus gold and silver. You set the spread before the run, and commission is charged on every trade. Swap is charged on every position held overnight.
The full metric set
Equity curve, profit factor, Sharpe ratio, maximum drawdown and the complete trade list.
Improve
The backtest comes back and Takumis reads it with you.
Takumis stops at the suggestion and waits for you. Nothing re-runs on its own. To let it keep going by itself, use Mega Mode.
It reads the backtest with you
It shows where the strategy lost money, which of your rules hurt the result, and one specific change to try next.
Testing on unseen data
A change that only fits the period you have been looking at is not worth much. Hold back part of the history and Takumis can test the change on data it was never tuned on.
Change a number, and only that number changes
Ask to change a number and Takumis changes exactly that number, then checks that the saved version matches what you asked for. If a rule itself changes, it tells you in the chat.
Mega Mode
Set a profit target, a drawdown limit and a budget, and Mega Mode runs the loop by itself: analyse, backtest, change, test again. It tries several versions at once on one part of the history, then picks the best one on a period it never saw.
Plan a single trade
One-off trade plans for a market event.
Workflows
Plan one trade around a market event. You confirm the plan, and Takumis builds it with its trigger inside, tests it on past data and gives it an expiry date, so a setup that made sense on Tuesday does not linger into Friday.
Publish and fork
Publish a strategy that passes the test, or copy someone else's and keep working on it.
The strategy gallery
Authors publish their own work, and a strategy is listed only if its backtest passes a mechanical qualification test. Passing asks you whether to publish, and you can decline. A listing goes stale once its backtest is more than a year old, until a fresh backtest re-qualifies it. Nothing in the gallery is a recommendation.
Fork anything published
Take any published strategy into your own private project and keep working on it with the AI.
Fork lineage
If your fork passes the test again and beats the best listed version by 15% or more on profit factor or payoff ratio (average win divided by average loss), it is listed beside the original.
Creator profiles
Every creator has a public profile with their published strategies and how often each one has been forked.
What we are building
None of this is in the beta yet. Each item moves up the page once it works.
So, what would you build?
Takumis is in private beta. Register to request access.