When you research with AI, it is easy to miss the moment to stop. Test one idea and you start wondering about another condition. Check that, and you want to try a different market. Ask AI, and there is usually a way forward. You can change the code and run it again.
I have spent about a year researching systematic trading this way.
I have enjoyed it. There is something satisfying about testing an idea that previously existed only in your head. If every question had first required me to work out whether I could implement it, I would probably have let many of them go. With AI, more things became possible to try.
But having more options came with a cost. I had to decide what not to try.
If I do not like a backtest result, for example, there is no shortage of things to adjust. I can change an entry condition, take a closer look at an unusually bad period, or apply the strategy to another market. AI can help with each attempt. A plausible explanation is usually available, too.
This is where research gets murky. Correcting a faulty assumption and changing conditions until a result looks appealing can feel remarkably similar while you are doing the work. Both involve editing code and running it again.
So the list of things to write down grew. Before an experiment: what am I trying to establish, and what result would make me drop the idea? Afterwards: which conditions did I change after seeing the output? Keeping these notes is a chore. Without them, though, I can later remember even my reasons for making changes as more convincing than they were.
Something similar happened with the code. Once a run finishes and produces numbers, it is tempting to relax. But numbers alone do not guarantee that the calculation is right. Signal timing and order timing might not line up. Rounding an order quantity might remove a trade I had intended to make. An equity curve does not readily reveal these details.
My requests to AI therefore had to become specific. Beyond asking it to implement a strategy, I needed to check why an order appeared on a particular date, which data it had read, and how the quantity had been calculated. A discussion of the overall results would eventually bring us back to one trade or one line of arithmetic.
From the outside, this work might not look like much. A whole day can end with the conclusion that a result needs checking again. But if I skip that check, I no longer know what I am relying on when I move to the next experiment.
AI helped quite a bit here. I could work through existing code with it, compare separate runs, and look for assumptions I had missed. It could not take responsibility for whether I was steering my questions towards the answer I wanted. That remained my job.
This is also why I want to start a blog. Research kept producing another task, leaving little time to explain the path I had taken. Writing it down should make it clearer which parts I understand and which parts I have glossed over.
I plan to take one experiment or one question at a time: why I thought it was worth trying, what turned out differently in practice, and what I changed next.
Looking at the notes from the past year, I do not expect to run out of things to write about soon. I am just a little behind on sorting them out.