The gap in how people learn
General tutorials teach concepts. They cannot tell you why your particular strategy behaves the way it does, so you are left guessing, re-reading and changing three things at once. That is slow, and it teaches the wrong lessons.
What an AI is good at here
- Explaining a result in plain words: why a run lost money, what a metric means, why a stop was hit.
- Turning an idea into explicit rules, which you then read and correct.
- Suggesting the next experiment, such as testing on other coins or splitting results by market condition.
- Spotting suspicious results: very few trades, one coin doing all the work, a curve that looks too smooth.
A learning loop that works
Change one thing at a time, and ask why before you change the next.
- State the hypothesis in a sentence: what should make this strategy earn money?
- Backtest it on several coins and a period it was not built around.
- Read the verdict and the regime table, then ask why for anything surprising.
- Change one thing, rerun, and compare.
- Validate on held-out data, then paper-trade.
Good questions to ask
- “Which trades did most of the damage, and what did they have in common?”
- “Does this only work in bull markets?”
- “If I halve the leverage, what happens to the worst drawdown?”
- “Is the evidence for this strong enough to paper-trade?”
An illustration
Imagine you ask: “Why did this strategy lose in March?” A useful answer reads the run, not the internet. It might say that most of the losing trades were longs entered in a sideways regime, that they were stopped out quickly and repeatedly, and that the strategy made money in the trending months around it. It would then suggest a test, such as adding a trend filter and rerunning on the same period. That is a hypothesis you can check in minutes, which is the whole point. (This is an invented example, to show the shape of a good answer.)
Where AI still needs checking
An assistant that guesses is worse than none. The value comes from being grounded in your real data: the Lab's Copilot explains runs by reading them with the same tools it uses to build and run strategies, and asks you before anything expensive or live. Still read the rules it writes, treat a confident explanation as a hypothesis to test, and remember that no assistant can promise a strategy will keep working. The same applies to your own agent. See bring your own AI.