Let your AI review the rules you write

More and more trading platforms advertise AI-enabled trading: an AI that watches the market and decides for you when to buy and sell. A fair question is whether Aurono will let an AI decide too.

Aurono will keep the buying and selling with your own rules. What the latest release of Aurono does add is all of your own data in a form your AI tool of choice can read, plus a starting prompt that asks the AI the question worth asking.

Why your rules keep the final say

Aurono rests on three beliefs, and two of the three settle this question. The first belief is that Aurono is locally controlled: decisions live on your device, you write the rules, and Aurono carries the rules out. The second belief is that Aurono is fully inspectable: every trade can be traced to the specific rule that fired it. When something happens in your strategy, you can open the audit trail and see which rule, which price and which check led to the trade.

An AI making the call breaks both beliefs. The decision would no longer be yours, and “the model thought so” is not a rule anyone can inspect afterwards. The same reasoning is behind writing your rules down before the market tests you: a rule is valuable because you chose the rule with a calm head, and the rule holds when the market does not.

The line holds for anything Aurono builds later too: an AI can suggest, and only you create or change a strategy.

What we built instead

An AI is still useful in a different role: as a second pair of eyes on the rules you wrote. Settings now has a section called “Data for AI analysis”. One click downloads everything Aurono has recorded about your strategies:

every strategy and every version of its parameters, with what each version earned
every completed round trip and every order, filled or not, simulated or live
every evaluation: what the price did, which rule fired, and why
what the price did a day after each decision
the same money in a simple buy-and-hold, for comparison
the full event log, for anyone who wants to check the figures

Every figure matches the figure the app shows. A README inside the download explains every column, so the AI does not have to guess what acb_price or BELOW_ACB means.

The download goes straight to your computer, and Aurono sends nothing to any AI service. Your exchange API keys and email settings are never in the file. Which AI you share the file with, and on what terms, is your decision.

Judge the decision, not the outcome

In an earlier post I wrote about Annie Duke’s warning against resulting: judging a decision by how the decision turned out instead of how the decision was made. A good rule can lose on an unlucky market, and a bad rule can win on a lucky one. Annie Duke reviewed the hands she won as carefully as the hands she lost.

Your data makes the same review possible. Aurono records every signal at the moment the signal happens, including the signals your rules held back: a sell that waited because the price was below your cost, a buy that waited because the cash was gone. Those held-back signals are the branches that did not happen, and Aurono writes them down before anyone knows how things turned out. The download adds where the price was a day later, so you can ask a sharper question than “did I make money”: did holding back protect me, or did holding back cost me?

CSV for AI chats, a database for AI code tools

You choose the format before you download:

Spreadsheet files (CSV) work with any AI chat, such as ChatGPT, Claude or Gemini, and open in Excel or Numbers. If your AI tool cannot open a ZIP file, unzip the download and upload the files inside.
Database (SQLite) is for AI tools that work with files on your computer, such as Claude Code. Instead of reading every row, the AI writes queries against the data, which suits long histories. The README includes a few example queries to start from.

There is one more option: you can leave out the evaluations where nothing happened, for AI tools that limit upload size. Leave the quiet evaluations in when you can. Quiet evaluations show how close each price move came to your triggers, which is exactly what the AI needs to judge whether a trigger is set too wide.

A prompt to start with

The README contains a suggested prompt, and after the download Settings shows the same prompt with a copy button. You can ask your own question instead, but the suggested prompt carries what we learned from testing, below. The prompt asks the AI to:

rank the strategies against buy and hold as a percentage of the money you put in, not in euros
review the winners as hard as the losers, and say whether the rules earned the result or something else did
for the losing strategies, tell apart triggers that never fired from triggers that fired and were held back, and check what the price did afterwards
suggest at most three changes, each with a result that would prove the change right or wrong
show the calculation behind every number, and give two confidence levels: one for the diagnosis, one for the suggested change

What we learned testing the export

Before releasing the export, we gave the download to ChatGPT, Gemini and Claude, compared their answers, and changed the export and the prompt after every round. A few lessons are worth knowing even if you write your own prompt.

Ask about results and you get resulting. Our first prompt asked which strategies did best and worst against buy and hold. The first answer ranked by euros and put a strategy at the top that had never placed a single trade: the strategy had kept its cash while its coin fell. The answer was arithmetically right and told us nothing about the rules. The prompt now asks the AI to separate what the rules did from what the market did, and the download includes the market’s part as a figure of its own.

A percentage needs the right base. Put €100 into a strategy, take €90 out again, and a €7 edge over buy and hold is 7% of what you put in, or 70% of the €10 that is left. One of the first analyses divided by what was left. The download now includes the total money put in, and the prompt asks for that comparison.

A win rate near 100% tells you nothing. Aurono does not sell below your average cost, so nearly every completed round trip is a winner by design. A weak strategy shows up elsewhere: as an unrealized loss, or as a rising market the strategy sat out. The README now explains this to the AI.

No evaluation means no trade. If your device is off or offline, your strategy is not evaluated, and a missed evaluation is also a missed trade. The money figures stay exact, but the figures belong to a rule that ran only part of the time. The download now shows what share of the expected evaluations each strategy recorded, so the AI can tell a rule that rarely fired from a rule that rarely ran.

The tools differ more than you might expect. The answers that ran code to analyse the files checked their own figures and found real patterns, including a winner whose profit came almost entirely from one short burst of trading. An answer that read the files as plain text attributed a total for all strategies to a single strategy, then gave that conclusion 98% confidence. The prompt now asks for code where possible, the calculation behind every number, which files could not be read in full, and confidence as high, medium or low instead of a precise-looking percentage.

An AI is a sharp colleague, and sometimes a confidently wrong one. The best answer we got corrected a sentence in our own README. The same answer also blamed a problem on a starting cost that, when we checked, was not the cause. Check any claim that would change your strategy against what the app itself shows.

Before you change anything

Nothing in the download can change a strategy: a downloaded file has no way back into Aurono. When the AI suggests a change, test the change first: run the new parameters in Lab → Simulate, or as a new strategy in shadow mode, and see whether the result the AI predicted actually shows up. Then decide, and make the change yourself.

Data for AI analysis is now available in the latest Aurono Start release. You will find it in Settings, below Data Management.