What is the biggest risk with AI-generated trading bots?

New comment
 

What is the biggest risk with AI-generated trading bots?

Bad logic, poor risk management, overfitting, or users not understanding the code?

 
David Chidiebere Chinweike:

What is the biggest risk with AI-generated trading bots?

Bad logic, poor risk management, overfitting, or users not understanding the code?

If your prompt is a blank slate, AI generates useless boilerplate/generic trading logic, e.g., an SMA(10), and SMA(20), and the rest is left to your "discretion."

AI can't analyze tick nor OHLC data, so good luck getting meaningful stoploss/risk parameters from AI.

Overfitting is a byproduct of machine learning or optimization. In the case of machine learning, you presumably asked AI for some form of it. In the case of optimization, AI can't backtest on tick nor OHLC data so it's all on you in your backtesting. In either case, the GIGO principle applies─garbage in, garbage out.

There are AI engines that generate detailed code comments which should, at least, help you understand the code.

 
Ryan L Johnson #:

If your prompt is a blank slate, AI generates useless boilerplate/generic trading logic, e.g., an SMA(10), and SMA(20), and the rest is left to your "discretion."

AI can't analyze tick or OHLC data, so good luck getting meaningful stoploss/risk parameters from AI.

Overfitting is a byproduct of machine learning or optimization. In the case of machine learning, you presumably asked AI for some form of it. In the case of optimization, AI can't backtest on tick or OHLC data so it's all on you in your backtesting. In either case, the GIGO principle applies─garbage in, garbage out.

There are AI engines that generate detailed code comments which should, at least, help you understand the code.

Yes, the risk is when you expect AI to do the thinking forgetting that AI is just an assistant. AI is not designed to build a profitable EA rather it can save you some time if used properly 
 
When there are too many interactions, she can very easily forget or mix them up.
In short, something will be missing.
 
Gerard William G J B M Dinh Sy #:
When there are too many interactions, she can very easily forget or mix them up.
In short, something will be missing.
That is an issue with the context window of the LLM. I switch to a new context window as soon as i hit 70% of the current context window. That way the LLM is able to maintain the scope of the work with guessing.
 
AI working, sometimes, as a black box could be the problem, working with the outcomes without understading what is happening is the problem, for me.
 
Isaac Uriel Arenas Caldera #:
AI working, sometimes, as a black box could be the problem, working with the outcomes without understading what is happening is the problem, for me.
Absolutely, AI as a tool to do something you have a solid idea about.
 
The biggest risk nobody talks about is overfitting. A bot can look perfect on backtests and fail completely in live markets because it was optimized for past data not future conditions. I learned this the hard way building and testing my own EAs on XAUUSD. The strategy that made 50% in backtests lost money live within weeks. The market conditions that trained it simply stopped repeating. Now I treat any backtest result with suspicion until it survives at least 3 different market conditions — trending, ranging, and high volatility.
 
just go with swing trades with small sl and huge tp with proper money management
 
Ridhdi Patel #:
just go with swing trades with small sl and huge tp with proper money management
I will test this out. I have a couple of questions which are:
  1. What do you look at for as a signal when swing trading?
  2. What metrics do you use to decide your SL?
 
Imoro Abdul-rahaman #:
The biggest risk nobody talks about is overfitting. A bot can look perfect on backtests and fail completely in live markets because it was optimized for past data not future conditions. I learned this the hard way building and testing my own EAs on XAUUSD. The strategy that made 50% in backtests lost money live within weeks. The market conditions that trained it simply stopped repeating. Now I treat any backtest result with suspicion until it survives at least 3 different market conditions — trending, ranging, and high volatility.
This is good to note. I usually  focus on forward testing rather than backtesting
New comment