Responses to Realistic Uses for AI in Trading
By Dave Mabe
Last week, I posted about realistic uses for AI in trading.
Several readers responded to share how they're using AI in their own trading.
Here's a summary of the best ones. (Names used with permission, lightly edited for clarity.)
Sebastian D.
So well said - computer-generated strategies have been around for a long time. Now, it's being democratized to some extent, but as Dennis Williams said, "We could post our trading rules on the front page of the Wall Street Journal and still people would not be able to make money from them."
Nico J.
One of the ways I use AI, as a new trader, is for research. I picked this up from something Michael said in Line Your Own Pockets.
Instead of having Claude create a strategy for me, I have it do a deep dive into ORBs, looking at all the possible permutations an ORB might take and the studies around it. I have it look at academic papers and elsewhere to provide the greatest background I can, and then I have it give me possible ways the trades might be automated.
After that, I make my own judgements, create an optimization in the Strategy Cruncher off a backtest, and get to work. The work it does is something I could never do myself—at least not without doing another PhD!
Matt H.
I'm using AI to code a stop loss based on R optimizer (runs across various ST_ columns and optimizes the multiplier of each at different aggression levels). I've also used it to make AFL's trade loop both record the progression of extremes and be able to trigger an exit type for an indicator. I'm just testing BB1min right now, but I also used AI to update the add_metrics.py script (with MabeKit) to easily add more (it has to update 4 different parts of the AFL code to do this, but I just run the script). The stop losses give significant improvement and confirm that, at least in my current strategy case, using a hard profit target is of limited value. HOWEVER, using a soft profit target (I'm not as skilled at sniping entries as you are yet) provides a significant improvement in win rate and total profit.
Anne O.
I'm learning every day. Right now, I'm working with Claude to create a central hub UI that allows me to link TradingView for scanning, Benzinga news feed, and Etrade to day trade. We're working on mostly non-strategy elements, but also presenting data in new ways, making it easier. As far as strategy goes, we review my trades together and build reports. I learned a lot in teaching my strategy to Claude; identifying the quantifiable elements and the discretionary bits became much clearer once Claude helped verbalize my primary strategy. As a result, we have found a few tweaks that are helping quantify some of those elements I previously thought were wholly discretionary. I am finding Claude to be a fantastic collaborator.
Another Brian
I built a template locally that is along the same lines as what you are describing.
The intent is not to have AI invent the strategy. The idea is to give it a defined process and then let AmiBroker do the work. So it is sort of an extension of “I use AI to develop my system,” but with the power of AmiBroker behind it.
I only let it go so far before I take over. It never comes up with the starter ideas. I still define the initial strategy and decide what is worth testing.
I used AI to help create three guideline documents:
a Dave guideline document
an AFL code guideline
an output guideline for how it presents information back to me
Then I used Codex locally on my PC to build a Python agent. The agent runs locally, but it uses the ChatGPT API to do the AI part.
That agent runs AmiBroker locally. It takes my starter strategy and runs it through the wringer, step by step, until it gets to the target-setting stage. That is where it stops.
It can run backtests, run explorations, test rules one at a time, create additional output columns, and add useful columns to a local library. It has already created a bunch of additional columns that I can use later.
I still have to check the important parts of the code, the column definitions, and the results. So it is not a black box. It is more like a structured assistant that does the heavy lifting and gives me better evidence to review.
Keeping everything local has been much better than using browser-based ChatGPT and downloading or uploading files. It makes fewer file-handling mistakes, and it is easier to keep the whole process organized.
Still experimenting and working out bugs, but the process seems to be there.
Great stuff here.
Notice the common theme: focused, specific workflows.
No grandiose "do everything for me" projects.
Pick out some of your trading grunt work and see if you can create an easier process for it.
A great place to start is your daily review.
Thanks to all who responded and shared!
-Dave
P.S. Do you wish you had a column library that would tell YOU how to make your strategy profitable? My column library is now included with MabeKit.